arXiv:math.GR/0605355 v1 13 May 2006
Axes in Outer Space Michael Handel and Lee Mosher May 13, 2006 Abstract We develop...
11 downloads
946 Views
903KB Size
Report
This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form
arXiv:math.GR/0605355 v1 13 May 2006
Axes in Outer Space Michael Handel and Lee Mosher May 13, 2006 Abstract We develop a notion of axis in the Culler–Vogtmann outer space Xr of a finite rank free group Fr , with respect to the action a nongeometric, fully irreducible outer automorphism φ. Unlike the situation of a loxodromic isometry acting on hyperbolic space, or a pseudo-Anosov mapping class acting on Teichm¨ uller space, Xr has no natural metric, and φ seems not to have a single natural axis. Instead our axes for φ, while not unique, fit into an “axis bundle” Aφ with nice topological properties: Aφ is a closed subset of Xr proper homotopy equivalent to a line, it is invariant under φ, the two ends of Aφ limit on the repeller and attractor of the source–sink action of φ on compactified outer space, and Aφ depends naturally on the repeller and attractor. We propose various definitions for Aφ , each motivated in different ways by train track theory or by properties of axes in Teichm¨ uller space, and we prove their equivalence.
Contents 1 Introduction 1.1 Characterizations of the axis bundle. 1.2 The main theorems. . . . . . . . . . 1.3 A question of Vogtmann. . . . . . . 1.4 Contents and proofs. . . . . . . . . . 1.5 Problems and questions . . . . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
3 4 8 9 9 12
2 Preliminaries 2.1 Outer space and outer automorphisms . . . . . . 2.2 Paths, Circuits and Edge Paths . . . . . . . . . . 2.3 Folds . . . . . . . . . . . . . . . . . . . . . . . . . 2.4 Train Track Maps . . . . . . . . . . . . . . . . . 2.5 The attracting tree T+ . . . . . . . . . . . . . . . 2.6 Geodesic laminations in trees and marked graphs 2.7 The expanding lamination Λ− . . . . . . . . . . . 2.8 Relating Λ− to T− and to T+ . . . . . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
15 15 20 21 24 28 33 35 38
1
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
3 The 3.1 3.2 3.3 3.4
ideal Whitehead graph Definition and structure of the ideal Whitehead graph Asymptotic leaves and the ideal Whitehead graph . . T+ and the ideal Whitehead graph . . . . . . . . . . . An example of an ideal Whitehead graph . . . . . . .
. . . .
40 41 43 44 45
4 Cutting and pasting local stable Whitehead graphs 4.1 Pasting local stable Whitehead graphs . . . . . . . . . . . . . . . . . . . . . . 4.2 Cutting local stable Whitehead graphs . . . . . . . . . . . . . . . . . . . . . . 4.3 The finest local decomposition . . . . . . . . . . . . . . . . . . . . . . . . . .
47 47 49 50
5 Weak train tracks 5.1 Local decomposition of the ideal Whitehead graph. 5.2 Folding up to a weak train track . . . . . . . . . . 5.3 Comparing train tracks to weak train tracks . . . . 5.4 Rigidity and irrigidity of Λ− isometries . . . . . . . 5.5 Examples of exceptional weak train tracks . . . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
52 53 54 56 58 61
6 Topology of the axis bundle 6.1 Continuity properties of the normalized axis bundle . . . 6.2 The Gromov topology on weak train tracks . . . . . . . . 6.3 Properness of the length map. . . . . . . . . . . . . . . . . 6.4 Applying Skora’s method to the Properness Theorem 6.1
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
65 65 67 70 73
7 Fold lines 7.1 Examples of fold paths . . . . . 7.2 Characterizing fold lines . . . . 7.3 Direct limits of fold rays . . . . 7.4 Legal laminations of split rays . 7.5 Weak train tracks on fold lines
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
81 81 83 85 88 92
. . . . .
. . . . .
. . . . .
2
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
1
Introduction
There are many interesting and fruitful analogies between the group of isometries of hyperbolic n-space Hn , the mapping class group of a finite type surface acting on Teichm¨ uller space, and the outer automorphism group Out(Fr ) of a rank r free group Fr acting on outer space Xr . In all cases the space has a natural compactification, obtained by adding a boundary at infinity to which the group action extends continuously. In all cases there is a class of “hyperbolic” elements whose extended action has source–sink dynamics, with a repeller and an attractor at infinity: loxodromic isometries of Hn ; pseudo-Anosov mapping classes of a surface [McC85]; and outer automorphisms of Fr which are fully irreducible [LL03], meaning that no proper, nontrivial free factor of Fr has a conjugacy class which is periodic under the action of the outer automorphism. In Hn and in Teichm¨ uller space, a “hyperbolic” element φ has a unique axis in the usual metric sense, a properly embedded, oriented, φ-invariant geodesic line which φ translates in the forward direction. The axis of φ depends only on the ordered pair (repeller, attractor), and this dependence is natural with respect to the group action: the axis is the unique geodesic line whose negative end converges to the repeller and whose positive end converges to the attractor. In Xr , however, it is not reasonable to expect such a nice notion of an axis for a fully irreducible φ ∈ Out(Fr ). There does not seem to be a natural candidate for a metric on Xr , and certainly not one that has the unique geodesic properties of Hn or Teichm¨ uller space. So among the many properly embedded lines in Xr whose two ends converge to the repeller and the attractor of φ, it seems hard to single out one of those lines in a natural way and call it the axis of φ. Nevertheless, among all such lines we shall select in a natural manner a particular collection of them, the union of which we call the “axis bundle” of φ, a subset Aφ ⊂ Xr with the following properties: Properness: Aφ is a closed subset of Xr proper homotopy equivalent to R. The negative end of Aφ limits in compactified outer space on the repeller, and the positive end limits on the attractor. Naturality: Aφ depends only on the ordered pair (repeller, attractor) in a manner which is natural with respect to the action of Out(Fr ). We shall do this only in the case where φ is nongeometric, meaning that it does not arise from a homeomorphism of a compact surface with boundary. The geometric case, while conceptually much simpler and more well understood (see [BH92], []), has some peculiarities whose inclusion in our theory would overburden an already well laden paper. Naturality means that for any nongeometric, fully irreducible φ, φ′ ∈ Out(Fr ) and any ψ ∈ Out(Fr ), if ψ takes the repeller–attractor pair of φ to the repeller–attractor pair of φ′ , then ψ takes Aφ to Aφ′ . One consequence of naturality is that Aφ is invariant under φ, because the repeller–attractor pair of φ is invariant under φ. Another is that axes are indifferent to iterates (positive powers) of φ: φn and φ have the same axis for all n > 0, because φn has the same repeller and the same attractor as φ. Certain differences between axes in hyperbolic space or Teichm¨ uller space on the one hand, and axis bundles in outer space on the other hand, may already be apparent from 3
the above features. First, the topological relation between the axis bundle and the real line is not homeomorphism but instead proper homotopy equivalence. Second, axes in Hn or Teichm¨ uller space are indifferent to inverses — the axis of φ−1 is the orientation reversal of the axis of φ. But as we have argued in [HM06b], it seems unnatural to insist that axes in Xr be indifferent to inverses, due to the phenomenon that a fully irreducible φ ∈ Out(Fr ) can have an expansion factor which is different than the expansion factor of its inverse. Although our paper [HM06a] offers a bare hint of an “indifference to inverses” property, by bounding the ratio of the logarithms of the expansion factors of φ and φ−1 , nevertheless we do not attempt here to construct a notion of axis which is indifferent to inverses. The main results of this work are a theorem which proves equivalence of several characterizations of the axis bundle — from some of which the naturality feature of the axis bundle is evident — and a theorem which proves the properness feature of the axis bundle. In the remainder of this introduction we shall describe the three characterizations of the axis bundle, and the methods of proof.
1.1
Characterizations of the axis bundle.
We shall characterize the axis bundle by presenting three different definitions, the equivalence of which is one of our major theorems. These definitions are motivated by various aspects of the theory of outer automorphisms, and by analogies with various properties of Teichm¨ uller geodesics. Fold lines. Fold lines form a natural class of paths in outer space arising from several concepts: train track maps due to Bestvina and Handel [BH92]; Stallings folds [Sta83]; and semiflow concepts behind Skora’s method for investigating homotopy properties of spaces of R-trees ([Sko], and see also [Whi93], [Cla05], [GL05]). Recall that elements of compactified outer space X r = Xr ∪ ∂Xr are represented by minimal, very small actions of Fr on R-trees, what we shall call “Fr -trees”. Two Fr -trees represent the same element of X r if and only if they are related by a homothety (a bijection that multiplies the metric by a constant) that is Fr -equivariant. An element of Xr itself is represented by a simplicial Fr -tree on which the action of Fr is free. Passing to the quotient modulo Fr of a free, simplicial Fr -tree, an element of outer space is also represented by a “marked graph”, a metric graph G without valence 1 vertices whose fundamental group is marked by an isomorphism Fr ≈ π1 (G); while we allow G to have valence 2 vertices, their presence or absence does not effect the element of Xr that G represents. The group Out(Fr ) acts on X r from the right: the image under φ ∈ Out(Fr ) of a marked graph G is a marked graph G · φ defined by precomposing the marking Fr ≈ π1 (G) with an automorphism of Fr that represents φ. Given a fully irreducible φ ∈ Xr , the repeller and attractor for φ are elements of ∂Xr represented by Fr -trees that we denote T− = T−φ and T+ = T+φ . A fold line in outer space is a continuous, injective, proper function R 7→ Xr defined by a continuous 1-parameter family of marked graphs t 7→ Gt for which there exists a family of homotopy equivalences hts : Gs → Gt defined for s ≤ t ∈ R, each of which preserves marking, such that the following hold:
4
Train track property: hts is a local isometry on each edge for all s ≤ t ∈ R. Semiflow property: hut ◦ hts = hus for all s ≤ t ≤ u ∈ R, and hss : Gs → Gs is the identity for all s ∈ R. Our first definition of the axis bundle is: Definition (1): Fold lines. Aφ is the union of the images of all fold lines γ : R → Xr such that γ(t) converges in X r to T−φ as t → −∞ and to T+φ as t → +∞. Notice that, just from the format of this definition, the axis bundle Aφ has the desired naturality feature, because the collection of fold lines in Xr is invariant under the action of Out(Fr ), and the points T−φ , T+φ ∈ ∂Xr depend naturally on φ. The construction of φ-periodic fold lines is given in detail in the beginning of Section 7, drawing on well known ideas reviewed in Section 2.2. Here is an outline of the construction. Recall that every fully irreducible φ ∈ Out(Fr ) is represented by an affine train track map, a homotopy equivalence g : G → G of a marked graph such that g takes vertices to vertices, g induces the outer automorphism φ on Fr ≈ π1 (G), and there exists λ > 1 so that the restriction of each iterate g i to the interior of each edge locally stretches the metric by the factor λi ([BH92], and see Section 2.4). The number λ = λ(φ) is independent of g and is called the expansion factor of φ. Let Gi = λ1i G · φi be renormalized marked graphs representing the orbit of G under powers of φ. For integers i ≤ j the iterate g j−i : G → G of the map g induces a marking preserving homotopy equivalence gji : Gi → Gj , and these maps satisfy the semiflow property, which is just a fancy way of saying that the following diagram commutes: ···
/ Gi
gi+1,i
/ Gi+1
/ ···
/ Gj−1 gj,j−1 6/ Gj
/ ···
gji
Moreover, the definition of train track maps and the renormalization of Gi together imply that gji satisfies the above train track property. Now one uses Stallings method of folds to factor the map g1,0 : G0 → G1 as a product of fold maps, and then one interpolates each fold map continuously to obtain a path in Xr from G0 to G1 . Concatenating the φi translates of this path for all i ∈ Z, one obtains a φ-periodic fold line in Xr . Source–sink dynamics on compact subsets of outer space [BFH97] implies that in X r this fold line converges in the negative direction to T− and in the positive direction to T+ . This shows that the point of Xr represented by G satisfies Definition (1) of Aφ . Notice that we do not claim that Aφ is the union of all φ-periodic fold lines. In fact, we do not even claim that Aφ is the union of all φi -periodic fold lines over all i ≥ 1. See Section 5.3 for an example of G ∈ Aφ that does lie on any φi -periodic fold line for any iterate φi . In particular, such a G is not the domain of a train track representative of any iterate φi , nor is G the domain of a nonsimplicial train track representative 1 of any iterate, a map that satisfies the definition of a train track representative except that it is not required to take vertices to vertices. 1 also
called an “efficient representative” in an unpublished preprint of T. White and in [LL04].
5
An analogy with Teichm¨ uller geodesics. Fold lines can also be motivated by an analogy with Teichm¨ uller geodesics, as follows. Recall that if S is a closed oriented surface of genus ≥ 2, each oriented geodesic γ in the Teichm¨ uller space T (S) has the following essentially unique description. Choosing a unit speed parameterization t 7→ σt of γ, there is a transverse pair of measured foliations F− , F+ with transverse measures denoted p |dy|, |dx|, respectively, such that for each t ∈ R the singular Euclidean metric dst = (et |dx|)R2 + (e−t |dy|)2 agrees with the conformal structure σt . If |dx|, |dy| are normalized so that S |dx| · |dy| = 1 then the pair F− , F+ depends up to joint isotopy only on the parameterization of γ. Also, in Thurston’s boundary ∂T (S) = PMF(S), the projective measured foliation space, the pair of points that is represented by F− , F+ depends only on γ as an oriented geodesic in T (S). Notice that the identity map on S underlies a system of maps σs 7→ σt that p satisfies the semiflow property. Consider now the renormalized metric ds′t = e−t dst = (|dx|)2 + (e−2t |dy|)2 , also in agreement with the conformal structure σt . For each leaf L of F− and each t ≤ u ∈ R, the restrictions to L of the metric ds′t and the metric ds′u are in agreement, because each agrees with the restriction of the transverse measure |dx|. To put it another way, each of the maps σt 7→ σu restricts to a local isometry on each leaf L of F− , with respect to the metrics ds′t and ds′u . Thus we have demonstrated an analogy with the maps Gs 7→ Gt in the definition of a fold line, each of which restricts to a local isometry on each edge of Gs . An even stronger analogy between Teichm¨ uller geodesics and fold lines will be used below to motivate Definition (2). Laminations, trees, and Λ− isometries. Thurston compactified Teichm¨ uller space with projective classes of measured laminations [FLP+ 79]. Skora proved a duality theorem between R-trees and measured foliations, which provides an alternative description of the boundary of Teichm¨ uller space in terms of R-trees [Sko90]. The boundary of outer space ∂Xr is described solely in terms of R-trees, no general theory of measured laminations having so far arisen. However, in [BFH97] Bestvina, Feighn, and Handel define the expanding lamination of a fully irreducible φ ∈ Out(Fr ) in terms of train track representatives of φ, and they prove a duality relation between the expanding lamination and T− which is a natural analogue of Skora’s duality. This relation is described by saying that the expanding lamination of φ is the “length zero lamination” of T− (see Section 3 of [BFH97], and Lemma 2.19 below). Because of this duality we shall denote the expanding lamination of φ as Λ− = Λφ− . The lamination Λ− can be regarded abstractly as an Fr -equivariant subset of ∂Fr × ∂Fr , or it may be realized as a family of leaves — bi-infinite geodesics — in any free, simplicial Fr -tree T ; see Sections 2.6 and 2.7 below. Given a free, simplicial Fr -tree T , an Fr -equivariant map fT : T → T+ is called a Λ− isometry if, for each leaf L of Λ− realized in T , the restriction of fT to L is an isometry onto a bi-infinite geodesic in T+ . Here is our next definition of the axis bundle: Definition (2): Supports of Λ− isometries to T+ . Aφ is the set of all elements of Xr represented by a free simplicial Fr tree T for which there exists a Λ− isometry fT : T → T + .
6
For example, given any train track representative g : G → G of φ, the universal cover e regarded as a free, simplicial Fr -tree, satisfies Definition (2). The details of why this T = G, is true are given in Corollary 2.20, but here is the idea. By lifting to the universal cover Ti of each marked graph Gi in the diagram above, we obtain Fr -equivariant maps g˜ji : Ti → Tj that satisfy the semiflow property for integer values i ≤ j, that is, the following diagram commutes for all i ≤ j: ···
/ Ti
g ˜i+1,i
/ Ti+1
/ ···
/ Tj−1 g˜j,j−1 7/ Tj
/ ···
g ˜ji
Each map g˜ji is an Fr -equivariant lift of the train track map g j−i : G → G, and the restriction of g˜ji to each edge of Ti is an isometry. In this situation, the trees Ti and the maps g˜ji form a direct system whose direct limit is an Fr tree representing T+ , and from the direct limit construction one obtains a map fi : Ti → T+ for each i. Moreover, from the construction of the expanding lamination Λ− it follows that for each leaf L of Λ− realized in each Ti , the restriction of gji to L is an isometry. Passing to the direct limit, it follows that the restriction of fi to L is an isometry, and so the map fi demonstrates that the point of Xr represented by Ti or by Gi satisfies Definition (2) of Aφ . Another analogy with Teichm¨ uller geodesics. Further motivation for Definition (2) comes again by an analogy with geodesics in Teichm¨ uller space T (S). Consider such a geodesic γ with all the associated data as above: σt , F− , F+ , |dy|, |dx|, dst , ds′t . Lifting to g is dual to an e the measured foliation Fe+ with transverse measure |dx| the universal cover S, R-tree T+ on which π1 (S) acts: as a point set T+ is obtained from Se by collapsing each leaf g as a pseudo-metric on of Fe+ to a point; as a metric space T+ is obtained by thinking of |dx| Se and passing to the associated metric space obtained by collapsing to a point each maximal subset of pseudo-metric diameter zero. Let f : Se → T+ be the collapse map. For each t, the conformal structure σt lifts to a π1 (S) equivariant conformal structure σ et , and we write the map f in the form ft : σ et → T+ . Under these conditions, the restriction of ft to each leaf L of Fe− realized in the singular Euclidean structure ds′t is an isometry, because the metric ds′t g restricted to L. Thus we can say that the restricted to L equals the transverse measure |dx| map ft is an F− -isometry (with respect to the metric ds′t ), and so each point σt on γ satisfies a property that is an analogue of Definition (2). To state this property precisely, given a geodesic γ determined by the data above, we say that σ ∈ T (S) satisfies property (2′ ) if and only if there exists a singular Euclidean structure ds agreeing with σ, and there is a map f : Se → T+ , such that locally f is modelled on the projection of a Euclidean rectangle to one e geodesic in S, e the restriction of f to L of its sides, and for each leaf L of F− realized as a ds is an isometry onto a bi-infinite geodesic in T+ . We have shown that each point on γ satisfies (2′ ), and the converse is easily seen to hold as well. Points of γ are thus characterized by property (2′ ), just as points in the axis bundle Aφ are characterized by (2). Train tracks. Here is our final definition of the axis bundle, which could have been stated first but has been relegated to the last. Given a fully irreducible φ ∈ Out(Fr ), while “train track representatives” of φ are well established objects of study, “train tracks” themselves 7
are not. We define a train track for φ to be a point of Xr that is represented by a marked graph G such that there exists a train track representative g : G → G of φ. Let TT(φ) ⊂ Xr be the set of train tracks of φ. Definition (3): Closure of train tracks. Aφ is the closure of
∞ [
TT(φi ) in Xr .
i=1
Remark. In Section 5.5 we produce an element of Aφ which is not a train track. We also believe there are easy examples showing that the closure of TT(φ) itself need not equal Aφ , although we have not produced such an example. This raises the question: does there N [ TT(φi )? exist some N ≥ 1 such that Aφ is the closure of i=1
1.2
The main theorems.
Having proposed three definitions of the axis bundle, our first main result shows that these three definitions are equivalent: Theorem 1.1. For each nongeometric, fully irreducible φ ∈ Out(Fr ) with repelling and attracting points T− , T+ ∈ ∂Xr and length zero laminations Λ− , Λ+ , definitions (1), (2), and (3) of Aφ are equivalent. In other words, for any G ∈ Xr the following are equivalent: (1) G lies on some fold line t 7→ Gt , t ∈ R, such that Gt → T− as t → −∞ and Gt → T+ as t → +∞. (2) There exists a Λ− isometry from the universal cover of G to T+ . (3) G is in the closure of the set of train tracks of iterates of φ. The set Aφ defined by this theorem is called the axis bundle of φ. While Definitions (1) and (3) may seem the most intuitive, Definition (2) occupies an important middle ground between intuition and rigor, being easier to work with when proving certain properties, in particular properness. For much of the paper we focus on Definition (2), and in Section 5 we introduce the terminology “weak train track” to refer to G satisfying Definition (2). Of course, once Theorem 1.1 is proved, then a weak train track is any element of the axis bundle Aφ . Note particularly that the natural dependence of Aφ on the ordered pair (T− , T+ ) ∈ ∂Xr × ∂Xr is an immediate consequence of definition (1) and also of definition (2). Our next main result establishes the properness feature of the axis bundle. Recall that a proper map f : A → B is a proper homotopy equivalence if there are proper maps f¯: B → A, ¯ : B × [0, 1] → A such that H is a homotopy between f¯ ◦ f and H : A × [0, 1] → B, and H ¯ is a homotopy between f ◦ f¯ and IdB . IdA , and H Theorem 1.2. If φ ∈ Out(Fr ) is nongeometric and fully irreducible, and if γ : R → Xr is a fold line such that γ(t) → T±φ as t → ±∞, then the inclusion map γ(R) ֒→ Aφ is a proper homotopy equivalence. Moreover, the end of Aφ corresponding to the negative end of L converges in X r to T− , and the end of Aφ corresponding to the positive end of L converges in X r to T+ . 8
1.3
A question of Vogtmann.
In [Vog02] Vogtmann asked: “Is the set of train tracks for an irreducible automorphism contractible?” We address this question only for the case of a nongeometric, fully irreducible φ ∈ Out(Fr ). The question requires some interpretation, because the notion of “train tracks” for φ is not well established in the literature. The notion of train tracks that we have defined above — points in outer space that support irreducible train track representatives of φ — is not appropriate, because the set of train tracks of φ has only countably many orbits under the action of the cyclic group hφi on Xr . By interpreting Vogtmann’s question instead using weak train tracks — elements of the axis bundle of φ — Theorem 1.2 gives a strong positive answer to her question, establishing not just the homotopy type but the property homotopy type of the set of weak train tracks Aφ . Relations to work of Los and Lustig. Another approach to Vogtman’s question is given in a preprint of Los and Lustig [LL04], who define a space of nonsimplicial train track representatives which they prove is contractible. One of our key technical results, Proposition 5.4, is very similar to Lemma 3.16 of [LL04] which also plays a key role in their results. Our results differ from theirs in several aspects. They study a space of maps between marked graphs representing elements of Xr , whereas our axis bundle is a subset of Xr ; see also the remarks at the end of Section 6.2. Their proof uses an interesting semiflow which is necessarily different from the semiflow used by Skora in [Sko], whereas we prove proper homotopy equivalence to the line by plugging directly into Skora’s theorem; see Section 6.4. Finally, in Section 5.5 we give an example of a weak train track that is not the domain of a nonsimplicial train track representative; hence, although the subset of Xr consisting of the domains of nonsimplicial train track representative can be shown to be a dense, contractible subset of the axis bundle, this subset is not equal to the whole axis bundle, and so it is not closed in Xr , and its proper homotopy type is not evident.
1.4
Contents and proofs.
First we describe the contents of the individual sections, then we prove the two main theorems by referring to results found in various sections. Section 2, Preliminaries. We give very detailed preliminaries, some of which is a compilation of previous work. Some details give new geometric perspectives of familiar objects such as train track representatives (Section 2.4) and laminations (Sections 2.6 and 2.7). Other details arise in making explicit what was only implicit in previous works, and in providing detailed proofs of results that, while previously evident, have not previously been recorded (see in particular Sections 2.5, Section 2.7, and Section 2.8). Section 3, The ideal Whitehead graph. In this section and the next we introduce the study of the ideal Whitehead graph of a nongeometric, fully irreducible φ ∈ Out(Fr ). Intuitively this graph describes topological and combinatorial aspects of the singularities of 9
φ. The graph has one edge for each singular leaf of the expanding lamination of φ, with two singular half-leaves impinging on the same vertex whenever those half-leaves are asymptotic. The analogous construction carried out for a singularity of the unstable foliation of a pseudoAnosov surface homeomorphism produces a circle with three or more vertices. In Out(Fr ) the topology of a component of the ideal Whitehead graph can be different from a circle, as we show by an example in Section 3.4. The ideal Whitehead graph of φ is defined in Section 3.1 in terms of the actions of automorphisms representing φ on the boundary of Fr . In terms of a train track representative g : G → G, we show that the ideal Whitehead graph of φ decomposes into the local stable Whitehead graphs of g, in a pattern dictated by the Nielsen paths of g. Sections 3.2 and 3.3 give other descriptions of the ideal Whitehead graph, in terms of: asymptotic relations among leaves of Λ− ; and branch points of the attracting Fr -tree T+ . Section 4, Cutting and pasting local stable Whitehead graphs. Given a nongeometric, fully irreducible φ ∈ Out(Fr ) and a train track representative g : G → G, we study how the decomposition of the ideal Whitehead graph of φ into local stable Whitehead graphs of g varies under some standard methods of varying g that were first described in [BH92]. This variation takes the form of cutting and pasting local stable Whitehead graphs, according to the formation or collapse of Nielsen paths of g. The main result of this section, Proposition 4.4, describes in a natural way those train track representatives which give the finest possible decomposition of the ideal Whitehead graph. Section 5, Weak train tracks. In this section we take up in earnest the study of the axis bundle, by proving several results about weak train tracks. We show that many structural features of train tracks, e.g. the induced local decomposition of the ideal Whitehead graph, have analogues for weak train tracks. A key technical result, Proposition 5.4, says roughly that if one fixes a train track G then for every weak train track H, one can obtain H from G by a sequence of Stallings folds provided two conditions hold: the local decomposition of H is no finer than the local decomposition of G; and the length of H is sufficiently small in terms of the geometry of G. As mentioned above, this result is closely related to a key result of [LL04]. Another important result in this section, Proposition 5.5, shows how to tell when a weak train track is a train track. As a corollary we obtain one of the implications of Theorem 1.1 by proving that a dense set of weak train tracks are train tracks. Also, in Section 5.4 we characterize when a weak train track H is the domain of a unique Λ− isometry, and more generally we describe the space of all Λ− isometries with domain H. Section 6, Topology of the axis bundle. This section proves several of our central results. Theorem 6.1 describes a proper homotopy equivalence, the length map, from the set of weak train tracks to the positive real axis (0, ∞). This theorem represents most of the work in proving the Properness Theorem 1.2. The proof is an application of Skora’s semiflow methods [Sko], which have also played a central role in other investigations of the homotopy types of spaces of R-trees [Whi93], [Cla05], [GL05]. We shall directly apply the main theorem of [Sko], for which we must set up the appropriate maps and verify the hypotheses of
10
that theorem. This is accomplished by applying the main results of the earlier sections, in particular Proposition 5.4 and the study of rigidity and irrigidity carried out in Section 5.4. Also proved in this section is the fact that the set of weak train tracks is a closed subset of outer space, a step in the proof of Theorem 1.1 (see below). Section 7, Fold Lines. In this section we initiate a detailed study of fold lines. The main result proved is the equivalence of Definition (1) of the axis bundle with the other definitions of the axis bundle. The hard work here is the proof that definition (1) implies definition (2): every point on a fold line is a weak train track. To do this, given a fold line t 7→ Gt in Xr , using only the assumption that limt→±∞ Gt = T± in X r we show that both the expanding lamination Λ− and the attracting tree T+ can be reconstructed by natural geometric processes stated in terms of the marked graphs Gt and the edge isometries hts : Gs → Gt . These processes allow us to construct a Λ− isometry from each Gt to T+ , proving that Gt is a weak train track. Along the way we also prove one other important piece in the proof of the Properness Theorem 1.2, that the length map restricted to each fold line is a homeomorphism between that fold line and the positive real axis. Throughout the paper we have probably erred on the side of more detail rather than less, particularly when the extra detail seems structurally interesting and can be illustrated with particular examples of train track maps; see for example Sections 3.4 and 5.4. We have tried to develop the theory of the axis bundle comprehensively, incorporating three different natural definitions into the statement of Theorem 1.1. However, we have isolated all the material regarding Definition (1) in Section 7, and so the reader who is willing to completely ignore Definition (1), and consider only those portions of the theory concerned with Definitions (2) and (3) will find the paper somewhat shortened. The Proof of Theorem 1.1. Let φ ∈ Out(Fr ) be nongeometric and fully irreducible, and denote T± = T±φ , Λ± = Λφ± . We use the terminology weak train track of φ to refer to an element of Xr satisfying Definition (2), being represented by a free, simplicial Fr -tree T for which there exists a Λ− isometry T 7→ T+ . The equivalence of definitions (1), (2), and (3) breaks into several implications. Definition (2) implies Definition (3): A dense set of weak train tracks are train tracks, as proved in Corollary 5.6. Definition (3) implies Definition (2): Every train track for φ is a weak train track, as proved in Lemma 2.20. The set of weak train tracks is closed, as proved in Lemma 6.2 (1). Together these prove that the closure of the set of train tracks is contained in the set of weak train tracks. Definition (2) implies Definition (1): Every weak train track of φ lies on a fold line for φ, as proved by a construction given in Proposition 7.1. Definition (1) implies Definition (2): Every point on a fold line for φ is a weak train track for φ, as proved in Theorem 7.2, which is stated in Section 7, and whose proof takes up the bulk of Section 7. ♦ The Proof of Theorem 1.2. Let A be the axis bundle of a nongeometric, fully irreducible φ ∈ Out(Fr ). In the beginning of Section 6 we define the length map Len : A → (0, +∞), 11
by normalizing each weak train track in a natural manner and taking the length of this normalization, and we prove that the length map satisfies the equivariance condition Len(T · φi ) = λ(φ)−i Len(T ) for any i ∈ Z Theorem 6.1 says that Len : A → (0, +∞) is a proper homotopy equivalence. Consider a fold line γ : R → Xr such that γ(t) converges in X r to T−φ as t → −∞ and to T+φ as t → +∞. From Theorem 1.1 it follows that γ(R) ⊂ A. Corollary 7.3 of Theorem 6.1 says that the restriction of the length map Len : A → (0, +∞) to γ(R) is a homeomorphism. It follows that the injection γ(R) ֒→ A is a proper homotopy equivalence. By combining the equivariance condition with the fact that Len : A → (0, +∞) is a proper homotopy equivalence, it follows that the set D = Len−1 [1, λ(φ)] is a compact fundamental domain for the action of the cyclic group hφi on A. Moreover, the end of A that corresponds i to the positive end of the fold line γ(R) is represented by the subset ∪+∞ i=1 φ (D). The unique limit point of this set in X r is T+ , because of the result from [BFH00] that the action of φi on X r converges uniformly on compact subsets of Xr to the constant map with value T+ , as i → +∞. Similarly, the end of A that corresponds to the negative end of γ(R) converges to T− . ♦
1.5
Problems and questions
Combinatorial structure of the axis bundle. Outer space has a combinatorial structure, a subdivision into closed subsets called “cells”, each parameterized by a simplex minus some subcomplex of that simplex, so that the intersection of any two cells is a common face; see [CV86] or Section 2.1 below. Question 1. Given a nongeometric, fully irreducible φ ∈ Out(Fr ), do the intersections of Aφ with the cells of outer space define a cell decomposition of Aφ ? The Properness Theorem 1.2 provides a proper homotopy equivalence from Aφ to (0, ∞) called the length function. If the previous question has an affirmative answer, perhaps level sets of the length function can be studied using the Morse theory ideas of Bestvina and Brady [BB97], in order to address the following: Question 2. Are the level sets of the length function contractible? Geometry of the axis bundles of φ and of φ−1 . The Properness Theorem 1.2 implies that Aφ has two ends, and that there exist subsets of Aφ that are compact and that separate the two ends; we call such a subset a cross section of Aφ . Our proof of the Properness Theorem carried out in Section 6 provides a construction of a cross section, using a Stallings fold argument. Fold ideas are also central to the techniques of [HM06a] relating the expansion factors of φ and of φ−1 , and we believe that these techniques may extend to relating their axis bundles as well. We propose here a question that would serve to explore this relation. To quantify the size of a cross section, fix an appropriate measurement of (pseudo)distance on Xr , for example: an Out(Fr )-equivariant geodesic metric for which closed balls are compact; or a combinatorial measurement where the distance between two points x, y ∈ Xr is 12
the minimal length of a chain of cells σ0 , σ1 , . . . , σL with x ∈ σ0 , y ∈ σL , and σi−1 ∩ σi 6= ∅ for all i = 1, . . . , L. Given a point x ∈ Aφ define the girth of Aφ at x to be the infimal diameter of a cross section of Aφ passing through x, and define the girth of Aφ to be the supremal girth over all points x ∈ Aφ . Given x ∈ Aφ ∪ Aφ−1 , define the joint girth at x of Aφ and Aφ−1 to be the infimal diameter of a compact subset of Xr which contains the point x and which contains cross-sections of Aφ and Aφ−1 . Define the joint girth of Aφ to be the supremum over all x ∈ Aφ ∪ Aφ−1 of the joint girth at x. There is no uniform bound for girth depending only on rank. The example in Section 7.1 of a nonconvergent split ray can be extended to a fold line for some reducible element φ ∈ Out(Fr ), and with some thought one sees that the union of fold lines for φ contains a plane properly embedded into Xr . By carefully interjecting a few more folds into the fold description of φ, one can produce nongeometric, fully irreducible elements of Out(Fr ) with axis bundles of arbitrarily large girth. Given a nongeometric, fully irreducible φ ∈ Out(Fr ), the ideas of [HM06a] give a correlation between the expansion factors of φ and of φ−1 , and they do so in a manner which, we believe, also gives a correlation between large girth of Aφ and some kind of quantitative measurement that tells us “how irreducible” φ is. We would therefore expect to obtain some control over the joint girths of the axis bundles of φ and φ−1 , as long as one of them is “very irreducible”. Problem 3. Fixing the rank r, give conditions on a nongeometric, fully irreducible φ ∈ Out(Fr ) which imply a bound on the girth of Aφ . Also, give conditions that would imply a bound on the joint girth of Aφ and Aφ−1 . Generalizations of axes. In an attempt to find a broader generalization of Teichm¨ uller geodesics, it would be interesting to obtain results not just about source sink pairs (T− , T+ ) for fully irreducible outer automorphisms, but for more arbitrary pairs of points in ∂Xr . Question 4. For what pairs of points (T0 , T1 ) in ∂Xr is there a good theory of axes, mimicking our results for source sink pairs (T− , T+ )? Singularity types. As mentioned above, in Section 3 we present the ideal Whitehead graph of a nongeometric, fully irreducible outer automorphism as a description of the singularity structure of that outer automorphism. First recall the singularity structure of a pseudo-Anosov homeomorphism f : S → S of a closed, oriented surface S. The index of a singularity s of f is defined to be i(s) = 1 − k2 where s has k unstable P directions and k stable directions. The Euler-Poincar´e formula gives an index equation s i(s) = χ(S), summed over all singularities of f . Consider now a nongeometric, fully irreducible φ ∈ Out(Fr ). If C is a component of the ideal Whitehead graph, define its index to be the number i(C) = 1 −
#vertices(C) 2
P The index of φ is defined to be i(φ) = C i(C) summed over all components of the ideal Whitehead graph (modulo the action of Fr ). In contrast to the situation on a surface, rather than an index equation there is in general only an index inequality i(φ) ≥ 1 − r: in [GJLL98] 13
the Fr -tree index i(T+φ ) is defined and is proved to satisfy i(T+φ ) ≥ 1 − r, and the results of Section 3 imply that i(φ) = i(T+φ ). In [BF] it is proved that the inequality i(φ) ≥ 1 − r is an equality if and only if T+φ is a geometric Fr -tree, which in our present context occurs if and only if φ is a parageometric outer automorphism. Question 5. What values of the index deficit i(φ)−(1−r) are possible? Does the maximum index deficit go to +∞ as r → +∞? As a finer invariant, define the index type of φ to be the list of indices of the components of the ideal Whitehead graph of φ, written in increasing order. For instance when r = 3 then the possible index types which sum to 1 − r = −2 are (− 21 , − 12 , − 21 , − 12 ), (− 12 , − 12 , −1), (− 21 , − 23 ), and (−1, −1), and the possible index types whose sum is strictly greater than −2 are (− 12 , − 21 , − 12 ), (− 21 , − 12 ), (− 21 ), and (−1). See Section 3.4 for an example in F3 where the index type is (−1). Masur and Smillie proved that on a finite type surface, all possible index types, which sum to the Euler characteristic, are acheived by pseudo-Anosov homeomorphisms, with a few exceptions in low genus [MS93]. Question 6. What possible index types, which satisfy the index inequality i(φ) ≥ 1 − r, are acheived by nongeometric, fully irreducible elements of Out(Fr )? Even finer than the index type is the isomorphism type of the ideal Whitehead graph of φ. In the analogous situation of a pseudo-Anosov surface homeomorphism f : S → S, a singularity s of f has an ideal Whitehead graph whose isomorphism type is determined by the index: the unstable directions at s are arranged in a circle, and so if the index equals i then the ideal Whitehead graph of s is just a circle with −1 − 2i vertices and edges. In contrast, the example given in Section 3.4 is of a nongeometric, fully irreducible φ ∈ Out(F3 ) with a connected ideal Whitehead graph which is an arc with 4 vertices and 3 edges. There should be other examples in F3 with a connected ideal Whitehead graph having 4 vertices which is not an arc. Question 7. What isomorphism types of graphs actually occur among the ideal Whitehead graphs of fully irreducible outer automorphisms of Fr ?
14
2 2.1
Preliminaries Outer space and outer automorphisms
Outer automorphisms. Once and for all fix the integer r ≥ 2. Let Fr = hx1 , . . . , xr i be the free group of rank r with free basis x1 , . . . , xr . Let Aut(Fr ) be the automorphism group. Let iγ : Fr → Fr denote the inner automorphism iγ (δ) = γδγ −1 . Let Inn(Fr ) = {iγ : γ ∈ Fr } be the inner automorphism group. Let Out(Fr ) = Aut(Fr )/ Inn(Fr ) be the outer automorphism group, which acts on conjugacy classes of elements and of subgroups of Fr . An outer automorphism φ ∈ Out(Fr ) is reducible if there exists a nontrivial free decomposition Fr = A1 ∗ . . . ∗ Ar ∗ B such that φ permutes the conjugacy classes of A1 , . . . , Ar . If φ is not reducible then it is irreducible. If φi is irreducible for all integers i ≥ 1 then φ is fully irreducible; this is also known in the literature as “IWIP”, an acronym for “irreducible with irreducible powers”. φ is geometric if there exists a homeomorphism f : S → S of a compact surface with boundary S, and an isomorphism Fr ≈ π1 (S) which conjugates φ the outer automorphism of π1 (S) induced by f . If φ is not geometric then it is nongeometric. Fix a standard rose Rr , a graph with one vertex v and directed edges e1 , . . . , er , and identify Fr = π1 (Rr , v) where xi is the class of the loop ei . This induces an isomorphism of Aut(Fr ) with the group of homotopy classes of homotopy equivalences of the pair (Rr , v), and an isomorphism of Out(Fr ) with the group of homotopy classes of homotopy equivalences of Rr . We use these isomorphisms without comment henceforth. Outer space. A graph is a finite 1-complex G in which every vertex v has valence(v) ≥ 2. The rank of G is the rank of π1 G. The valence condition implies that every homotopy equivalence between graphs is surjective. In a graph of rank r, an Euler characteristic P argument shows that v 21 (valence(v) − 2) = r − 1, and so there are at most 2r − 2 vertices of valence ≥ 3. Also, valence(v) ≤ 2r for each v, and if in addition valence(v) ≥ 3 for each vertex v then G has at most 3r − 3 edges. A metric graph is a graph G equipped with a geodesic metric. A similarity or homothety between two metric graphs G, G′ is a homeomorphism h : G → G′ such that d(hx, hy) = λ d(x, y) for some constant λ > 0 called the stretch factor of h. We use the notation Len(·) to represent length of objects in a metric graph G. For example: Len G is the sum of the length of the edges of G; for a path p : [a, b] → G we use Len p for the path length; and if c represents a homotopy class of maps S 1 → G then we use Len c to represent the minimal length of a representative of c. Sometimes to clarify the context we use LenG (·) to mean length in the graph G. A marked graph is a metric graph G equipped with a homotopy equivalence, usually denoted ρ = ρG : Rr → G, called a marking of G. Each marking ρG induces an isomorphism Fr ≈ π1 (G, ρG (v)), and this sets up a bijection between the set of markings of G modulo homotopy and the set of isomorphisms Fr ≈ π1 (G) modulo inner automorphism; we use this correspondence henceforth without comment. Given two marked graphs G, G′ , a homotopy equivalence f : G → G′ preserves marking if f ◦ ρG is homotopic to ρG′ . Two marked graphs G, G′ are (isometrically) equivalent if there exists an isometry G 7→ G′ that preserves 15
marking, and G, G′ are projectively equivalent if there exists a similarity G 7→ G′ that preserves marking. Culler and Vogtmann’s outer space Xr is the set of projective equivalence classes of marked graphs [CV86]. Define Xbr to be the set of (isometric) equivalence classes of marked graphs. By our conventions, a marked graph may contain valence 2 vertices, and may be subdivided or unsubdivided at such points without altering the equivalence class in Xr or in Xbr . With this flexibility, our definition of Xr is equivalent to that of [CV86] where a marked graph has no valence 2 vertices. We shall review below the topologies on Xr and Xbr , after discussing the interpretation of these spaces in terms of R-trees. Notation: Often we blur the distinction between a marked graph G, its isometric equivalence class in Xbr , and even its projective equivalence class in Xr . When we need more formality, we use (ρG , G) as notation for a marked graph, and [ρG , G] for its equivalence class, projective or nonprojective as the context makes clear. The group Out(Fr ) acts on the right of Xr and Xbr by changing the marking: given a marked graph (ρG , G), and given φ ∈ Out(Fr ) corresponding to a homotopy equivalence h : Rr → Rr , the action is defined by [ρG , G] · φ = [ρG ◦ h, G] Any homotopy equivalence g : G → G of a marked graph (ρG , G) determines a homotopy equivalence h = ρ¯G ◦g ◦ρG : Rr → Rr well-defined up to homotopy independent of the choices of ρG and its homotopy inverse ρ¯G within their homotopy classes, and so g determines an outer automorphism φ of Fr ; we say that g represents φ. We will always assume that g is an immersion on edges. If in addition g takes vertices to vertices then we say that that g is a topological representative of φ. Note that [ρG , G] · φ = [ρG ◦ h, G] = [g ◦ ρG , G] Actions on trees. We review group actions on R-trees, focussing on the group Fr . An R-tree is a metric space T such that for any x 6= y ∈ T there exists a unique embedded topological arc, denoted [x, y] ⊂ T , with endpoints x, y, and this arc is isometric to the interval [0, d(x, y)] ⊂ R. Given x ∈ T , two arcs [x, y], [x, y ′ ] determine the same germ or the same direction at x if [x, y] ∩ [x, y ′ ] 6= {x}. This is an equivalence relation on arcs with endpoint x, and the set of equivalence classes is called the set of germs or directions at x. A triod based at a point x ∈ T is a union of three arcs with endpoints at x no two of which have the same germ at x. The valence of x is the cardinality of the set of directions, equivalently, the number of components of T − x. The valence is ≥ 3 if and only if there exists a triod based at x. We say that T is simplicial if the points of valence 6= 2 form a discrete set. If T is simplicial and valence is everywhere finite then T is homeomorphic to a CW-complex. Every isometry f : T → T of an R-tree is either elliptic meaning that f has a fixed point, or hyperbolic meaning that f has an axis ℓ, an f -invariant isometrically embedded copy of the real line on which the translation length inf x∈T d(x, f (x)) is minimized. In the elliptic case the translation length equals zero. 16
Consider an isometric action of Fr on an R-tree T , denoted A : Fr → Isom(T ). Given γ ∈ Fr and x ∈ T we use various notations including tγ for A(γ) and γ ·x for tγ (x) = A(γ)(x). For a subset S ⊂ T we denote γ · S = {γ · x x ∈ S}. The action A is minimal if there does not exist a nonempty, proper, Fr -invariant subtree; this implies that T has no valence 1 points, because the valence 1 points are Fr -invariant and their complement is a nonempty Fr -invariant subtree. The action A is very small if the stabilizer of every nondegenerate arc is either trivial or a cyclic group generated by a primitive element of Fr , and the stabilizer of every triod is trivial. We shall refer to a very small, minimal action of Fr on an R-tree T as an Fr -tree. To each Fr -tree T we associate the length function ℓT ∈ RC , where ℓT [γ] is the translation length of the action of tγ on T . Consider two Fr -trees T, T ′ . A map f : T → T ′ is equivariant if γ · f (x) = f (γ · x) for x ∈ T . We say that T, T ′ are equivariantly isometric or isometrically conjugate if there is an equivariant isometry T 7→ T ′ , and they are topologically conjugate if there exists an equivariant homeomorphism. A result of Culler and Morgan [CM87] says that T, T ′ are isometrically conjugate if and only if ℓT = ℓT ′ in RC . Outer space in terms of trees. Given a marked graph (ρG , G) with universal covering e → G, we always assume that a basepoint in the universal cover G e has been chosen map pr : G e which projects to ρG (v), thereby inducing an action of Fr on G by covering translations. This e a free, simplicial Fr -tree whose isometric conjugacy class is well-defined independent makes G of choices. Conversely, every free, simplicial Fr -tree T determines a quotient marked graph GT = T /Fr , whose marking has homotopy class depending only on the isometric conjugacy e = TG , clearly we have TGT = T and GTG = G, and so we obtain class of T . Denoting G a natural bijection between Xˆr and the isometric conjugacy classes of free, simplicial Fr e then ℓT = ℓG in RC . trees. This bijection agrees with the length map to RC : if T = G Under this bijection the right action of Out(Fr ) on Xˆr can be described as follows: given a free, simplicial Fr tree T with action A : Fr → Isom(T ), and given φ ∈ Out(Fr ), choose Φ ∈ Aut(Fr ) representing φ, and let T · φ be the Fr tree with underlying R-tree T and with action A ◦ Φ : Fr → Isom(T ); this action is well-defined up to conjugacy independent of the choice of Φ. We can therefore describe Xbr either as marked graphs up to marked isometry or as free, simplicial Fr -trees up to isometric conjugacy; outer space Xr has a similar pair of descriptions. Given a free, simplicial Fr -tree T , depending on context we use [T ] for the point in Xr or Xbr represented by T , but we very often abuse notation by deleting the brackets and letting T stand for the Fr -tree as well as the corresponding point in Xr or Xbr . Ends, automorphisms, and lifts. The Gromov boundary of the free group Fr is a Cantor set denoted ∂Fr . The action of Aut(Fr ) on Fr extends to an action on ∂Fr ; the extension b : ∂Fr → ∂Fr . of Φ ∈ Aut(Fr ) is denoted Φ Given a free simplicial Fr -tree T , the space of ends of T is denoted ∂T , and there is a natural homeomorphic identification ∂Fr ≈ ∂T which we will use without further comment. Given γ ∈ Fr , the map tγ : T → T extends to a map of ends denoted tˆγ : ∂T → ∂T , whose fixed point set is equal to the pair of endpoints t± γ of the axis of tγ .
17
Suppose that G is marked graph and g : G → G is a homotopy equivalence representing e will be denoted g˘ : G e → G, e where the “\breve” φ ∈ Out(Fr ). A lift of g to the Fr -tree G notation ˘ is meant to be a reminder that the lifted map is not Fr -equivariant (unless φ is the identity and g preserves marking). The map g˘ extends to a map of ends denoted e → ∂ G. e There is a bijection between the set of lifts g˘ : G e →G e of g and the set of gˆ : ∂ G automorphisms Φ : Fr → Fr representing φ. This bijection is characterized by the twisted equivariance equation g˘ ◦ tγ = tΦ(γ) ◦ g˘ for all γ ∈ Fr . (2.1) The bijection is also characterized by the condition that g˘ and Φ extend to the same homeˆ : ∂Fr → ∂Fr . A simple computation using Equation 2.1 shows that if omorphism gˆ = Φ ′ e e g˘, g˘ : G → G are the lifts of g with corresponding automorphisms Φ, Φ′ ∈ Aut(Fr ) that represent φ, and if γ ∈ Fr is the unique element such that Φ′ = iγ ◦ Φ, then g˘′ = tγ ◦ g˘. ˆ are used interchangeably, depending on context. Notation: The notations gˆ and Φ The topology of outer space. The topology on Xr has three equivalent descriptions which we review: the cellular topology; the length or axes topology; and the Gromov topology. We use all three of these descriptions at various places in the paper. The cellular topology [CV86]. Consider a marked graph G. With respect to the natural vertex set consisting of points of valence ≥ 3, denote the edge set by E(G). The open cone of G is the set of elements in Xbr represented by marked graphs G′ for which there exists a homeomorphism G 7→ G′ that preserves marking; the open cell of G in Xr is similarly defined. Note that if g : G → G is a homeomorphism of a marked graph that preserves marking then g takes each edge to itself. It follows that if G′ is in the open cone of Xbr containing G then any two homeomorphisms G 7→ G′ that preserve marking induce the same bijection E(G) ↔ E(G′ ), and therefore the lengths of edges on G′ determine a well-defined vector L(G′ ) ∈ [0, ∞)E(G) . The map L is a bijection between the open cone containing G and the set (0, ∞)E(G) . By projectivizing, we obtain a bijection between the open cell of G in Xr and the interior of the simplex P[0, ∞)E(G) . The closed cone in Xbr containing G is the set of elements represented by marked graphs G′ for which there exists a quotient map p : G → G′ that preserves marking and that collapses to a point each component of a subforest F ⊂ G; the closed cell in Xr containing G is similarly defined. The map p induces a well-defined bijection E(G) − E(F ) ↔ E(G′ ), and therefore there is a well-defined vector LG′ ∈ [0, ∞)E(G) which assigns zero to edges in F and assigns the length of the corresponding edge of G′ to each edge in G − F . Again L is a bijection between the closed cone containing G and a certain subset of [0, ∞)E(G) , namely those vectors which assign coordinate zero to a subset of E(G) that forms a subforest of G. The closed cone of G is thus identified with a subset of [0, ∞)E(G) which is the union of the interior and certain of the faces. By projectivizing, we obtain a bijection between the closed cell of G in Xr and a union of faces of the simplex P[0, ∞)E(G) . Note that if F ⊂ F ′ are faces of the simplex and F is a face of the closed cone of G then F ′ is also a face of the closed cone of G.
18
The cellular topology on Xbr is the weakest topology such that for each marked graph G, the closed cone containing G is a closed subset of Xbr , and the map L from this cone to [0, ∞)E(G) is a homeomorphism onto its image. The cellular topology on Xr is the quotient by projectivization of the cellular topology on Xbr . We need some facts about the cellular topology. First, the closed cone (simplex) of G′ is a face of the closed cone (simplex) of G if and only if there is a quotient map G 7→ G′ that preserves marking and that collapses to a point each component of some subforest of G. Second, the closed cells (simplices) are locally finite: for any G′ there are only finitely many closed cells (simplices) that contain G′ , and hence by the weak topology there is a neighborhood of G′ that intersects only those finitely many cells (simplices) that contain G′ . Given a marked graph G define its length ℓ(G) to be the sum of the lengths of its edges. This gives a well-defined function on Xbr which, on each closed cell, is just the sum of the coordinates. It follows that, in terms of the cellular topology: Proposition 2.1. The map ℓ : Xbr → (0, ∞) is continuous. ♦
The length or axes topology [CM87]. Let C be the set of nontrivial conjugacy classes in Fr , give RC the product topology and let P RC be the associated projective space. For each marked graph G, each c ∈ C corresponds bijectively, via the marking ρG , with an immersion S 1 → G also denoted c, which is well-defined up to a homeomorphism of the domain, and whose length LenG (c) is well-defined. The map G → LenG defines an injection Xbr ֒→ RC and an induced injection Xr ֒→ P RC . These injections may also be defined in terms of translations lengths of free, simplicial Fr -trees representing elements of Xr . We give Xr and Xbr the topologies that make these injections homeomorphisms onto their images; we shall occasionally refer to these as the length topologies on Xr and Xbr . It follows that the natural action of R+ on Xbr , where t · G is obtained from G by multiplying the metric by the constant factor t, is an action by homeomorphisms with quotient space Xr . The Gromov topology [Pau89]. This topology on Xbr is defined by the following basis. Given an Fr -tree T ∈ Xbr , ǫ > 0, a finite subset K ⊂ T , and a finite subset P ⊂ Fr , define the basis element V (T ; ǫ, K, P ) to be the set of Fr -trees T ′ ∈ Xbr for which there exists a finite subset K ′ ⊂ T ′ and a relation R ⊂ K × K ′ such that R projects onto K and onto K ′ , and the following hold: R is an ǫ-almost isometry: For all x, y ∈ K and x′ , y ′ ∈ K ′ , if (x, x′ ), (y, y ′ ) ∈ R then |d(x, y) − d(x′ , y ′ )| < ǫ. R is P -equivariant: For all x ∈ K, x′ ∈ K ′ , and α ∈ P , if α · x ∈ K and (x, x′ ) ∈ R then α · x′ ∈ K ′ and (α · x, α · x′ ) ∈ R.
In [Pau89] the Gromov topology is defined by forming a basis as above but using compact subsets of T and T ′ instead of finite subsets; as shown in Proposition 4.1 of [Pau89] the definition given above is a basis for the same topology. The quotient topology on Xr is also called the Gromov topology. The following theorem is a combination of results of [CV86] and [Pau89]: Theorem 2.2. The cellular topology, the length topology, and the Gromov topology are identical on Xbr and on Xr . ♦ 19
Compactification and boundary of outer space. The closure of Xr in P RC is denoted b be the X r . Its boundary is ∂Xr = X r − Xr . Both X r and ∂Xr are compact [CM87]. Let X r C C inverse image of X r under the projection R − {O} 7→ P R , equivalently, the closure of Xbr b − Xb , which is just the inverse image under projection of ∂X . in RC − {O}. Let ∂ Xbr = X r r r We next relate compactified outer space to actions on trees. b there exists an F -tree T such Theorem 2.3. (Culler–Morgan [CM87]) For each ℓ ∈ X r r that ℓ = ℓT , and T depends uniquely on ℓ up to equivariant isometry. Extending our notation, when ℓ = ℓT as in this theorem, we use [T ] to denote the b or X as the context makes clear. corresponding point in X r r
b and given T with ℓ = ℓ, the following are equivalent: Lemma 2.4. Given ℓ ∈ X r T (1) ℓ ∈ Xbr .
(2) ℓ(c) is bounded away from zero for c ∈ C. (3) The action of Fr on T is free and properly discontinuous. Proof. It is obvious that the first item implies the second. Since the action is by isometries the second item implies the third. Finally, the third item implies that the quotient map from T to G := T /Fr is a covering space. Since T is an Fr -tree, G is a finite graph and so T is simplicial. ♦
2.2
Paths, Circuits and Edge Paths
e → G. A proper Consider a marked graph G and the universal covering projection pr : G e e if it is map σ ˜ : J → G with domain a (possibly infinite) interval J will be called a path in G an embedding or if J is finite and the image is a single point; in the latter case we say that e is homotopic rel endpoints to a σ ˜ is a trivial path. If J is finite, then every map σ ˜:J →G unique (possibly trivial) path [˜ σ ]; we say that [˜ σ ] is obtained from σ ˜ by tightening. e that differ only by an orientation preserving We will not distinguish between paths in G change of parametrization. Thus we are interested in the oriented image of σ ˜ and not σ ˜ itself. If the domain of σ ˜ is finite, then the image of σ ˜ has a natural decomposition as a e1 · E e2 · . . . · E ek , where E ei for 1 < i < k is an edge of G, e E e1 is the terminal concatenation E ek is the initial segment of some edge. If the endpoints of the segment of some edge, and E e1 and E ek are full edges. The sequence E e1 · E e2 · . . . · E ek is image of σ ˜ are vertices, then E called the edge path associated to σ ˜ , and the number k is called the combinatorial length of this edge path. When the interval of domain is half-infinite then an edge path has the e1 · E e2 · . . . or . . . · E e−2 · E e−1 , and in the bi-infinite case an edge path has the form form E e−1 · E e0 · E e1 · E e2 · . . .. ...·E e Thus a map A path in G is the composition of the projection map pr with a path in G. σ : J → G with domain a (possibly infinite) interval will be called a path if it is an immersion or if J is finite and the image is a single point; paths of the latter type are said to be trivial. If J is finite, then every map σ : J → G is homotopic rel endpoints to a unique (possibly trivial) path [σ]; we say that [σ] is obtained from σ by tightening. For any lift σ ˜ :J →Γ 20
of σ, [σ] = pr[˜ σ ]. We do not distinguish between paths in G that differ by an orientation preserving change of parametrization. The edge path associated to σ is the projected image of the edge path associated to a lift σ ˜ . Thus the edge path associated to a path with finite domain has the form E1 · E2 · . . . · Ek−1 · Ek where Ei , 1 < i < k, is an edge of G, E1 is the terminal segment of an edge, Ek is the initial segment of an edge; the combinatorial length again equals k. We reserve the word circuit for an immersion σ : S 1 → G. Any homotopically non-trivial map σ : S 1 → G is homotopic to a unique circuit [σ]. As was the case with paths, we do not distinguish between circuits that differ only by an orientation preserving change in parametrization and we identify a circuit σ with a cyclically ordered edge path E1 E2 . . . Ek , and again the combinatorial length equals k. ˜:H e →G e is a lift then If h : H → G is a homotopy equivalence of marked graphs and if h ˜ ˜ e for each path σ ˜ in H we denote [h(˜ σ )] by h# (˜ σ ), and for each path or circuit σ in H we denote [h(σ)] by h# (σ). Given a self homotopy equivalence g : G → G of a marked graph and a path or circuit σ in G, a decomposition σ = . . . · σi · σi+1 · . . . is called a splitting if k k k g# (σ) = . . . · g# (σi ) · g# (σi+1 ) · . . . for all k ≥ 0; in other words, to tighten g k (σ) one need only tighten the g k (σi ). Throughout this paper we will identify paths and circuits with their associated edge paths. A path or circuit crosses or contains an edge if that edge occurs in the associated edge path. For any path σ in G define σ ¯ to be ‘σ with its orientation reversed’. For notational simplicity, we sometimes refer to the inverse of σ ˜ by σ ˜ −1 . Bounded cancellation. The bounded cancellation lemma was introduced in [Coo87] and generalized to the form we need in [BFH97]. Let Lip(f ) be the lipschitz constant of a map f. Lemma 2.5. [Bounded Cancellation, Lemma 3.1 of [BFH97]] Let T0 be a free, simplicial Fr tree, let T be an Fr -tree representing a point in X r , and let f : T0 → T be an Fr -equivariant map. There exists a constant BCC(f ) ≤ Lip(f ) · Len(T0 /Fr ) such that for each geodesic segment α ⊂ T0 , the image f (α) is contained in the BCC(f ) neighborhood of the straightened image f# (α).
2.3
Folds
In this subsection we define morphisms, edge isometries, and fold paths. We also define fold maps and Stallings fold sequences, and we describe relations between these objects and fold paths. Turns and gates. In any graph a turn is an unordered pair of directions with a common base point. The turn is nondegenerate if it is defined by two distinct directions and is degenerate otherwise. Consider two graphs T, T ′ and a map H : T → T ′ which is locally injective on each edge. For each x ∈ T there is an induced map Dx H from the set of directions of T at x to the set of directions of T ′ at H(x): given a direction d at x represented by an oriented path α with
21
initial endpoint x, Dx H(d) is defined to be the direction at H(x) represented by H(α); the map Dx H is well-defined, independent of the choice of α representing d. We use DH either as an abbreviation for Dx H when x is understood, or as the union of all of the maps Dx H over all x ∈ T . Given an Fr -equivariant map of Fr -trees H : T → T ′ , we define the gates at a point x ∈ T with respect to H to be the equivalence classes of directions at x where d, d′ are equivalent if Dx H(d) = Dx H(d′ ). In applications where the base point x of a direction is a vertex we can identify directions with oriented edges E in the link of x, and if in addition H takes vertices to vertices then Dx H(E) is the first edge in the edge path associated to H(E). We have here extended the terminology of a “gate” which in the literature has so far been applied only to train track maps; see Section 2.4. Piecewise isometries and edge isometries. An Fr -equivariant map f : T → T ′ from a free, simplicial Fr -tree T to an Fr tree T ′ is a piecewise isometry 2 if each edge of T may be subdivided into finitely many segments on each of which f is an isometry. We usually use the even stronger property that f restricted to each edge of T is an isometry, to which we refer by saying that f is an edge isometry or that f is edge isometric. For a piecewise isometry to be an edge isometry is equivalent to it being locally injective on each open edge. When T ′ is also free and simplicial then we also define a piecewise isometry or edge isometry of marked graphs G = T /Fr 7→ T ′ /Fr = G′ , meaning that the lifted map T → T ′ is a piecewise isometry or edge isometry. Equivalently, an edge isometry between marked graphs is any homotopy equivalence that preserves marking and restricts to a local isometry on each open edge. Fold paths. A fold path in outer space is a continuous, proper injection γ : I → Xr , where I ⊂ R is connected, so that there exists a continuous lift γ b : I → Xbr , a marked graph Gr representing γ b(r) for each r ∈ I, and maps gsr : Gr → Gs for each r ≤ s in I, satisfying the following: (1) Train track property: Each gsr is an edge isometry. (2) Semiflow property: gtr = gts ◦ gsr for each r ≤ s ≤ t in I, and gss : Gs → Gs is the identity for each s ∈ I. A fold path γ whose domain I is noncompact on both ends is called a fold line. Fold paths are investigated in detail in Section 7. For now we turn to well known constructions that lead to simple examples of fold paths. Fold maps. Consider a marked graph G and two distinct oriented edges e, e′ of G that have the same initial vertex v. Parameterize e, e′ by arc length from their initial vertex, using the notation e(·), e′ (·). Choose a number S such that e([0, S])∩e′ ([0, S]) = {v}. For s ∈ [0, S] define Gs to be the graph obtained from G by identifying e(t) to e′ (t) for each t ∈ [0, s]. The quotient map G 7→ Gs is a homotopy equivalence, because of the fact that e(s) 6= e′ (s). Push forward the metric on edges of G by the quotient map G 7→ Gs , defining a metric on edges 2 A “piecewise isometry” is the same thing as a “morphism”, as commonly defined in the literature, but restricted so that the domain is free and simplicial.
22
of Gs which extends naturally to a geodesic metric on all of Gs . Postcompose the marking of G by the homotopy equivalence G 7→ Gs , defining a marking of Gs . It follows that the quotient map G 7→ Gs preserves the marking and is an edge isometry; when s = 0 this is just the identity map on G. When s > 0 we say that Gs is obtained from G by a length s fold of e and e′ , and the quotient map G 7→ Gs is called a fold map from G to Gs . Consider the map γ b : [0, S] → Xbr defined by γ b(s) = Gs . First, b γ is a continuous path in Xbr from G to GS , and so it descends to a continuous path γ : [0, S] → Xr which, its domain being compact, is proper. Second, for 0 ≤ s ≤ t ≤ S the induced map fts : Gs 7→ Gt is a fold map. Third, the family of maps fts satisfies the train track property and the semiflow property, by construction. It follows that γ is a fold path. Either of the paths γ : [0, S] → Xr or γ b : [0, S] → Xbr , or any of their orientation preserving reparameterizations, are called interpolating paths or just interpolations of the fold map G 7→ GS . Stallings fold sequences. Given an edge isometry h : G → H between marked graphs, a Stallings fold sequence for h is a finite sequence of edge isometric fold maps G = G0
f1,0
/ G1
f2,1
/ ···
fJ,J−1
/ 3 GJ = H
(2.2)
h
whose composition equals h. If we define fji for integer values 0 ≤ i ≤ j ≤ J to be the composition fj,j−1 ◦ · · · ◦ fi+1,i : Gi → Gj , then the maps fji are edge isometries satisfying the semiflow property fki = fkj ◦ fji for integer values 0 ≤ i ≤ j ≤ k ≤ J. For j = 1, . . . , J let [j − 1, j] 7→ Xbr be an interpolating path for the fold map fj,j−1 : Gj−1 → Gj , and extend the map fj,j−1 to a family of edge isometries fts : Gs → Gt , j − 1 ≤ s ≤ t ≤ j that satisfies the semiflow property. Concatenating the paths [j − 1, j] 7→ Xbr for j = 1, . . . , J defines a continuous path b γ : [0, J] 7→ Xbr , and the entire family of edge isometries fts defined for 0 ≤ s ≤ t ≤ J satisfies the semiflow property. Thus b γ descends to a continuous path γ : [0, J] → Xr which is a fold path, called the Stallings fold path of the Stallings fold sequence (2.2). To complete the construction of Stallings fold paths, we review Stallings original topological construction of fold sequences [Sta83], and adapt this construction to our present metric setting where fold maps are required to be edge isometric. At the same time we shall use this construction to bound the length of the fold sequence. In a Stallings fold sequence (2.2), the fold fj+1,j : Gj → Gj+1 is said to be complete if it satisfies the following: assuming fj+1,j to be a length S fold of oriented edges e, e′ of Gj , then S is equal to the maximum value of s such that restrictions of the map fJ,j : Gj → GJ = H to e([0, s]) and to e′ ([0, s]) are the same path in H. In other words, fj+1,j is complete if the maximal possible initial segments of e, e′ are folded immediately. Define the combinatorial length of an edge isometry h : G → H as follows. While at first regarding G and H with their natural vertex sets consisting solely of vertices of valence ≥ 3, subdivide H at the images of the vertices of G, which adds at most 2r − 2 new valence 2 vertices to H and at most 2r − 2 new edges, bringing the number of edges in H up to at most 5r − 5. For each edge e of G we may regard h(e) as an edge path in the subdivided graph H, and we define the combinatorial length of h to be the sum of the combinatorial lengths of h(e) over all edges e of G. Equivalently, having subdivided H, if we then subdivide G at the 23
inverse image of the vertex set of H, and if we refer to the edges of this subdivision as the edgelets of G with respect to h, then the combinatorial length of h is the number of edgelets of G with respect to h. Lemma 2.6. Each edge isometry h : G → H which is not an isometry has a sequence of complete Stallings folds. The length of any such fold sequence is bounded above by the combinatorial length of h. Proof. Subdivide H at the images of the vertices of G. Since the map h is a local injection on each edge and the graphs G, H have no valence 1 vertices, if h were locally injective at each vertex then there would be an extension of h to a covering map h′ : G′ → H where the graph G′ deformation retracts to G (see [Sta83]). Since h is a homotopy equivalence it would follow that h = h′ is a homeomorphism and therefore h is an isometry, contrary to the hypothesis. There exist therefore two oriented edges e, e′ of G with the same initial vertex v, and there exists a positive number R ≤ min{Len(e), Len(e′ )}, such that h maps e([0, R]) and e′ ([0, R]) to the same oriented path in H. The maximal such R has the property that h′
h(e(R)) = h(e′ (R)) is a vertex of H and e([0, R]) ∩ e′ ([0, R]) = {v}. Define G −→ G′ to be h′
h′′
the length R fold of e and e′ , and so h factors as G −→ G′ −−→ H. Noting that each of e([0, R]) and e′ ([0, R]) is a union of edgelets of G with respect to h, identifying e([0, R]) with e′ ([0, R]) has the effect of identifying their edgelets in pairs, and so the number of edgelets of G′ with respect to the edge isometry h′′ — the combinatorial length of h′′ — is strictly less than the number of edgelets of G with respect to h — the combinatorial length of h. The existence of a sequence of complete Stallings folds now follows by induction on combinatorial length. The same argument shows that for any sequence of complete Stallings folds (2.2), the combinatorial length of fJ,j : Gj → GJ = H is a strictly decreasing function of j, proving the desired bound. ♦
2.4
Train Track Maps
A topological representative g : G → G of an outer automorphism φ is an irreducible train track map if g k E is an immersion for each edge E and each k > 0 and if for every pair of edges E and E ′ there exists l > 0 such that g l (E) crosses E ′ . Theorem 1.7 of [BH92] states that every irreducible φ ∈ Out(Fr ) is represented by an irreducible train track map. Moreover, there exists λ > 1, called the expansion factor for φ, there exists a metric on G, and there exists a homotopy of g so that, still using g for the homotoped map, for all edges E the restriction g E uniformly expands length by the factor λ. We will always assume that the graph G is equipped with such a metric, and we occasionally emphasize this point by referring to g as an affine train track map. A nondegenerate turn in G is illegal with respect to g : G → G if its image under some iterate of Dg is degenerate, and otherwise the turn is legal. A path σ = E1 · E2 · . . . · Ek−1 · Ek ¯i , Ei+1 } for 0 ≤ i ≤ k − 1. A path σ ⊂ G is legal if it contains or takes each of the turns {E contains only legal turns. It is immediate from the definitions that Dg maps legal turns to legal turns and that g maps legal paths to legal paths. Note that g expands the length of any legal subpath by the factor λ. 24
We define the gates of a train track map g : G → G as follows. The map g induces a map Dg on the set of directions at vertices of g. The gates at a vertex x with respect to H are the equivalence classes of directions at x where d, d′ are equivalent if there exists i ≥ 1 such that (Dg)i (d) = (Dg)i (d′ ). It follows that two directions are in the same gate if and only if the turn they make is illegal. Note the distinction between gates of a train track map and gates of an Fr -equivariant map of Fr -trees which is defined in Section 2.3. Local Whitehead Graphs. Consider an irreducible train track representative g : G → G of an irreducible φ ∈ Out(Fr ). For each point x ∈ G, the local Whitehead graph at x, denoted W (x) or more explicitly W (x; G), is the graph whose vertices are the directions di at x and with an edge connecting di to dj if there exists an edge E and k > 0 such that the path g k (E) takes the turn (di , dj ). Since φ is irreducible, W (x) is connected (see section 2 of [BFH97]). For each x ∈ G, the map Dx g extends to a simplicial map W (x) 7→ W (g(x)) also denoted Dx g, because each turn at x taken by some g k (E) maps to a turn at g(x) taken by g k+1 (E). If x ∈ Per(g), then the local stable Whitehead graph at x, denoted SW (x) or SW (x; G), is the subgraph of W (x) obtained by restricting to the periodic directions at x and the edges between them. Since some iterate g k induces a continuous map Dx g k from W (x) onto SW (x), e and it follows that SW (x) is connected. There are analogous definitions of W (˜ x) = W (˜ x; G) e e SW (˜ x) = SW (˜ x; G) for any vertex x˜ of G; these can also be defined as the lifts of W (x) and SW (x). It follows that W (˜ x) and SW (˜ x) are connected. p Nielsen Paths. A path σ is a periodic Nielsen path if g# (σ) = σ for some minimal p ≥ 1 called the period of σ. A Nielsen path is a periodic Nielsen path with period one. A periodic Nielsen path is indivisible if it can not be written as a non-trivial concatenation of periodic Nielsen subpaths. A path σ is a periodic pre-Nielsen path if it is not a periodic Nielsen path k (σ) is a periodic Nielsen path. but some g# We record information about Nielsen paths in the following lemma.
Lemma 2.7. Assume that g : G → G is an irreducible train track map. (1) There are only finitely many indivisible periodic or pre-periodic Nielsen paths for g. (2) An indivisible periodic or pre-periodic Nielsen path σ decomposes into subpaths σ = αβ¯ where α and β are legal and the turn (¯ α, β) is illegal. Moreover, if αn and βn are the terminal segments of α and β that are identified by g n , then the αn β¯n ’s are an increasing sequence of subpaths of σ whose union is the interior of σ. (3) For every path or circuit σ there exists K(σ) ≥ 0 such that for all k ≥ K(σ), there is a k (σ) into subpaths that are either legal or are indivisible periodic Nielsen splitting of g# paths. (4) There exists K ∗ , L∗ > 0 so that if σ is a path that decomposes as the concatenation of two legal subpaths each of length at least L∗ then K(σ) ≤ K ∗ . (5) If σ is a bi-infinite path with exactly one illegal turn then one of the following holds for all k ≥ K ∗ .
25
k • g# (σ) is legal. k • g# (σ) splits as Rk · τk · Rk′ where Rk and Rk′ are legal rays and τk is an indivisible periodic Nielsen path.
Proof. The first three items follow from Lemma 4.2.5 and Lemma 4.2.6 of [BFH00]. For the fourth item, let L∗ = Len(G) + min{Len(e) e is an edge of G}
If σ is legal then K(σ) = 1 so we may assume that σ has exactly one illegal turn. Decompose σ as σ = σ1 σ ¯2 where σ1 and σ2 are legal and let σ1,n and σ2,n be the terminal segments of σ1 and σ2 that are identified by g n . If there does not exist n > 0 such that σi,k = σi,n for all k ≥ n then the closure of ∪n σ1,n σ ¯2,n is an indivisible Nielsen or pre-Nielsen path by Lemma 4.2.5 of [BFH00] and σ splits into an initial legal path, followed by ∪n σ1,n σ ¯2,n followed by a terminal legal path. By item (1), the pre-Nielsen paths become periodic Nielsen paths after a uniformly bounded number of iterations, and so K(σ) is uniformly bounded. We may therefore assume that there is such an n. By Lemma 2.5 (see also Lemma 3.9 of [BH92]) the length of σi,n is bounded above by Len(G). Removing an initial segment of σi , i = 1, 2, that is disjoint from σi,n does not change K(σ). Since Len(σi ) ≥ L∗ we may therefore remove such initial segments so that σ has endpoints at vertices and has length at most 2 Len(G) + 2 max{Len(e) e is an edge of G}. There are only finitely many such σ and we take K ∗ to be the maximum of their K(σ)’s. This completes the proof of (4). Lemma 4.2.2 of [BFH00] implies that a bi-infinite line with exactly one illegal turn splits as R · τ · R′ where R and R′ are legal rays and τ decomposes as the concatenation of two legal subpaths each of length at least L∗ . The fifth item therefore follows from the third and fourth items and the fact that g# maps legal paths to legal paths and indivisible periodic Nielsen paths to indivisible periodic Nielsen paths. ♦ Assumption: We assume from now on that for any irreducible train track map g : G → G, the endpoints of indivisible periodic Nielsen paths are vertices. This can always be arranged by a finite subdivision. Principal vertices, principal lifts and rotationless train track maps. We now recall from [FH04] the concepts of principal automorphisms, principal lifts of train track representatives, and rotationless outer automorphisms. While we apply these concepts here only to fully irreducible outer automorphisms, more general definitions are given in [FH04] which apply to arbitrary outer automorphisms and their relative train track representatives. Let φ ∈ Out(Fr ) be fully irreducible, and consider Φ ∈ Aut(Fr ) representing φ and its ˆ : ∂Fr → ∂Fr . Denote the fixed point set of Φ ˆ by Fix(Φ), ˆ and denote boundary extension Φ ˆ ˆ the subset of nonrepelling fixed points of Fix(Φ) by FixN (Φ). We say that Φ is a principal ˆ contains at least three points. The set of principal automorphisms automorphism if FixN (Φ) representing φ is denoted PA(φ). Two elements Φ, Φ′ ∈ P A(φ) are equivalent 3 if there exists γ ∈ Fn such that Φ′ = iγ ◦ Φ ◦ i−1 γ . If the fixed subgroup Fix(Φ) is non-trivial, then Fix(Φ) = hγi is infinite cyclic. This happens if and only if φ is geometric, in which case φ is represented by a pseudo-Anosov 3 This
equivalence relation is called “isogredience” in [Nie86] and [LL00a].
26
homeomorphism of a once punctured surface whose boundary is represented by γ (see Theorem 4.1 of [BH92]). The following lemma is contained in Lemma 2.3 and Lemma 2.4 of [BFH04] and in Proposition 1.1 of [GJLL98]. Lemma 2.8. Assume that Φ ∈ PA(φ). ˆ is a finite set of attractors. (1) If Fix(Φ) is trivial then FixN (Φ) ˆ is the union of t± with a finite set of (2) If Fix(Φ) = hγi is infinite cyclic, then FixN (Φ) γ tγ -orbits of attractors. − ˆ is an attractor then it is not t+ (3) If P ∈ FixN (Φ) γ or tγ for any nontrivial covering translation tγ .
ˆ 1 ) and Fix(Φ ˆ 2 ) are disjoint. Corollary 2.9. If Φ1 6= Φ2 ∈ PA(φ) then Fix(Φ ˆ 1 ) ∩ Fix(Φ ˆ 2 ). Observe that Φ−1 Φ2 is a covering translation Proof. Suppose that P ∈ Fix(Φ 1 tγ , for a nontrivial γ ∈ Fr . It follows that tγ (P ) = P and hence that γ ∈ Fix(Φ1 ) ∩ Fix(Φ2 ). In this case, φ is geometric and the corollary is well known; see for example Lemma 4.2 of [HT85]. ♦ Let φ ∈ Out(Fr ) be fully irreducible. We say that φ is rotationless if for each k > 0 and each principal automorphism Φk representing φk there is a principal automorphism Φ ˆ = FixN (Φ ˆ k ). In other words, the representing φ such that Φk = Φk and such that FixN (Φ) set of principal automorphisms and their nonrepelling fixed points in ∂Fr is stable under iteration. e →G e is a Let g : G → G be an irreducible train track representative of φ. A lift g˘ : G principal lift if the corresponding automorphism is a principal automorphism, and so two principal lifts g˘, g˘′ are equivalent if and only if g˘′ = tγ ◦ g˘ ◦ t−1 γ for some γ ∈ Fr . Consider a periodic vertex v ∈ G. Since v has at least one legal turn, v has at least two periodic directions. We say that v is a principal vertex if there are at least three periodic directions at v or if v is the endpoint of an indivisible periodic Nielsen path. The set of principal vertices is non-empty by Lemma 3.18 of [FH04]. If every principal vertex v ∈ G is fixed and if every periodic direction at a principal vertex is fixed then we say that g : G → G is a rotationless train track map. These definitions are connected by Proposition 3.24 of [FH04], which we recall. Proposition 2.10. Assume that g : G → G is an irreducible train track map for a fully irreducible φ ∈ Out(Fr ). Then φ is rotationless if and only if g : G → G is rotationless. Remark. It is clear that g p , and hence φp , is rotationless for some p ≥ 1. Some of our results are stated for rotationless φ and g : G → G; to apply these in general one replaces φ and g by suitable iterates. A pair of points x, y ∈ Fix(g) are Nielsen equivalent if they are the endpoints of a Nielsen path in G. The elements of a Nielsen equivalence class are usually referred to simply as a Nielsen class. It is well known, and easy to check, that x and y are Nielsen equivalent if and 27
e→G e and points x only if there is a lift g˘ : G ˜, y˜ ∈ Fix(˘ g ) that project to x and y respectively. In particular, for any lift g˘, the image of Fix(˘ g ) in G is either empty or is an entire Nielsen class. Moreover, if Fix(˘ g1 ) and Fix(˘ g2 ) are non-empty, then g˘1 and g˘2 are equivalent, as defined above in the context of principal lifts, if and only if Fix(˘ g1 ) and Fix(˘ g2 ) project to the same Nielsen class in Fix(g). It is useful to note that if Fix(˘ g1 ) ∩ Fix(˘ g2 ) 6= ∅ then g˘1 = g˘2 . This follows from the fact that g˘1 and g˘2 differ by a covering translation and the fact that every non-trivial covering translation is fixed point free. e of principal vertices v ∈ G as principal vertices It will be convenient to refer to lifts v˜ ∈ G e of G. Lemma 2.11. Assume that g : G → G is rotationless. Then PA(φ) 6= ∅. Moreover, g˘ is a principal lift if and only if Fix(˘ g ) 6= ∅ and some, and hence every, element of Fix(˘ g ) is a principal vertex.
Proof. The second statement follows from Corollary 3.14 and Corollary 3.22 of [FH04]. The first follows from the second and Lemma 3.18 of [FH04]; see also Proposition I.5 of [LL00b]. ♦ We also recall Lemma 3.23 of [FH04]. Lemma 2.12. If g is rotationless then every periodic Nielsen path has period one.
2.5
♦
The attracting tree T+
Given a fully irreducible φ ∈ Out(Fr ) there exists a unique point T+ = T+φ ∈ ∂Xr which is the attracting point for every forward orbit of φ in Xr ; this is proved in [BFH00]. We review here the construction and properties of the Fr -tree T+ . Although we shall not need this fact here, we mention the result of [LL03] that the action of φ on X r has source-sink dynamics −1 with sink T+φ and source T−φ = T+φ . Let g : G → G be an irreducible train track map that represents φ and let ρG : Rr → G be the marking on G. For all i ≥ 0, let Gi be the marked graph obtained from (g i ρG , G) by multiplying the metric by 1/λi , and so ℓGi (c) = ℓG (φi (c))/λi
for all c ∈ C and all i ≥ 0.
The map g : G → G induces a homotopy equivalence denoted gi+1,i : Gi → Gi+1 that preserves marking. By composition, for 0 ≤ i < j, the map g j−i : G → G induces the marking preserving homotopy equivalence gji = gj,j−1 ◦ · · · ◦ gi+1,i : Gi → Gj (see Figure 1). We fix the base point of Gi so that the marking g i ρG : Rr → Gi preserves base point. It follows that the maps gji : Gi → Gj preserve base point. b to an element ℓ ∈ X b whose projective class is φThe sequence ℓ converges in X Gi
r
+
r
invariant. To describe ℓ+ explicitly, consider c ∈ C represented by a circuit σ ⊂ G. If K(σ) is k as in Lemma 2.7 and k > K(σ) then g# (σ) splits as a concatenation of legal subpaths µi,k and P periodic indivisible Nielsen paths νi,k . Excluding the νi,k , define ℓ+ (c) = ( LG (µi,k ))/λk . Since g expands length of legal paths by the factor λ and there are only finitely many periodic indivisible Nielsen paths, ℓ+ (c) is independent of k > K(σ) and is the limit of the sequence ℓGi (c). It is immediate from the definition that ℓ+ (φ(c)) = λℓ+ (c) and that if c is represented by a legal circuit in G then ℓ+ (c) = ℓG (c). The Fr tree T+ that realizes ℓ+ is well defined up to isometric conjugacy. 28
gji
Gi
gi+1,i
/ Gi+1
gi+2,i+1
/ ···
gj,j−1
& / Gj
Figure 1: Notation associated to a train track map g : G → G. Direct limits. It seems to be well known that T+ is the Fr -equivariant Gromov limit of e i ; see e.g. [BF]. We need some slightly stronger facts about this convergence. the trees G Direct limits exist in the category of metric spaces and distance nonincreasing maps. That is, consider a direct system consisting of a partially ordered set I, an indexed collection of metric spaces (Ai )i∈I with metrics di , and an indexed collection of distance nonincreasing maps (fji : Ai → Aj )i<j , such that I is a directed set meaning that for all i, j ∈ I there exists k ∈ I with i ≤ k, j ≤ k, and the maps satisfy the compatibility property fkj ◦ fji = fki for i < j < k in I. Then there exists a metric space A∞ and distance nonincreasing maps f∞,i : Ai → A∞ satisfying the following universality property: given a metric space B and a system of distance nonincreasing maps hi : Ai → B compatible with the maps fji , meaning that hj ◦ fji = hi , there exists a unique distance nonincreasing map h∞ : A∞ → B such that h∞ ◦ f∞,i = hi : Ai → B for all i ∈ I. To construct A∞ , first let Aˆ∞ be the direct limit of the system (Ai ), (fji ) in the category of sets [Spa81]: points of A∞ are equivalence classes of the disjoint union of the Ai where x ∈ Ai is equivalent to y ∈ Aj if there exists k with i, j ≤ k such that fki (x) = fkj (y). Then define a pseudometric on Aˆ∞ where for any two equivalence classes [x], [y] we have d∞ ([x], [y]) = inf dk (fki (x), fkj (y)) k≥i,j
The direct limit A∞ is the natural quotient metric space of Aˆ∞ , where [x], [y] ∈ A∞ are equivalent if and only if d∞ ([x], [y]) = 0. The map f∞,i : Ai → A∞ is the composition of the quotient maps Ai → Aˆ∞ → A∞ . ei . We lift the base point of Gi to G e i , and we choose lifts e ei → Consider the trees G gi+1,i : G e i+1 of gi+1,i : Gi → Gi+1 for all i, so that the maps gei+1,i preserve base points, thereby G ei into an Fr tree, and making each map e making G gi+1,i into an Fr equivariant map which ei → G ej is an isometry when restricted to each edge. We obtain a direct system with e gji : G defined inductively by e gji = e gj,j−1 ◦e gj−1,i . Whenever Φ ∈ Aut(Fr ) representing φ is implicitly ei → G ei for all i ≥ 0 (see chosen, the associated lifts of gi : Gi → Gi shall be denoted g˘i : G Figure 2). These maps satisfy the following properties: • g˘i restricted to each edge is a homothety with stretch equal to λ. • geji g˘i = g˘j e gji .
• (Twisted equivariance, see (2.1)) g˘i ◦ tγ = tΦ(γ) ◦ g˘i for each γ ∈ Fr .
While we recall that the notation g˘ is meant to suggest twisted equivariance (2.1), on the other hand the notation ge is meant to suggest Fr -equivariance. 29
fg =fg0 fg1 fg2
e e=G G V 0 g ˘=˘ g0
g ˜1,0
/G e1 V
g ˜2,1
g ˘1
/G e2 V
g ˜3,2
/ ...
g ˘2
T+ W
Φ+
Figure 2: Notation associated to a train track map g : G → G and automorphism Φ representing φ. Maps denoted g˜ depend only on g and are Fr -equivariant. Maps denoted g˘ depend on g and on Φ and satisfy the twisted equivariance equation (2.1). The maps fgi are defined after Theorem 2.13, and Φ+ is defined in Theorem 2.15. The following theorem and corollary seem to be folklore. See e.g. [HM06b] for a proof of Theorem 2.13, and see Section 7.3 below for a proof of a more general fact in the nongeometric case. ej → G ei , in the category of metric Theorem 2.13. The direct limit of the direct system e gij : G spaces with an Fr action whose morphisms are Fr equivariant, distance nonincreasing maps, is an Fr tree in the same homothety class as T+ . ♦ In the context of this theorem, we shall normalize the metric on T+ so that T+ is identified e → T+ , and with the direct limit. From Theorem 2.13 we obtain a direct limit map fg : G e by construction fg is an isometry when restricted to each edge of G. As the notation fg suggests, this map is well-defined up to isometric conjugacy of the Fr action on T+ (see Figure 2), depending only on g : G → G, and not depending on the other choices made in ej → G ei . This well-definedness is a consequence the definition of the direct system e gij : G of the direct limit description of T+ , from which it follows that T+ is the metric space e0 where the distance between x e0 equals the limit of associated to a pseudometric on G ˜, y˜ ∈ G k k the nonincreasing sequence Len g# (pr[˜ x, y˜]) λ . Clearly the pseudometric depends only on g and therefore so does T+ . e is a simplicial Fr -tree Corollary 2.14. Assume that g : G → G is rotationless and that G obtained from the universal cover of G by choosing a marking and a lifting of the base point. e → T+ such that for all x e letting Then there is a surjective equivariant map fg : G ˜, y˜ ∈ G, σ = pr[˜ x, y˜], the following are equivalent. (1) fg (˜ x) = fg (˜ y ). k (2) For some k ≥ 0, the path g# (σ) is either trivial or Nielsen.
In particular, fg restricts to an isometry on all legal paths. e=G e0 and let fg : G e → T+ be the direct Proof. Assume the notation of Theorem 2.13 with G limit map. 30
k We first show that (2) implies (1). By construction, g# (σ) is trivial if and only if gek,0 (˜ x) = gk,0 (˜ e y ) and the latter implies that fg (˜ x) = fg (˜ y ). To prove (1) assuming that σ is a Nielsen or pre-Nielsen path, it suffices to assume that σ is indivisible. By Lemma 2.7-(2), there exist x ˜i → x ˜ and y˜i → y˜ such that gek,0 (˜ xi ) = e gk,0 (˜ yi ) for all i and all sufficiently large k. It follows that fg (˜ xi ) = fg (˜ yi ) for all i and hence that fg (˜ x) = fg (˜ y ). k Suppose now that fg (˜ x) = fg (˜ y ). If each g# (σ) is non-trivial, then Lemma 2.7 implies k that g# (σ) splits as a concatenation of legal paths and indivisible Nielsen paths for all sufficiently large k. If µ is any legal path in this splitting, then the distance between e gk,0 (˜ x) and e gk,0 (˜ y ) is uniformly bounded below by the length of µ in contradiction to the definition k of fg and the assumption that fg (˜ x) = fg (˜ y ). Thus no such µ exists and g# (σ) is Nielsen. We have now proven the equivalence of (1) and (2). Since a legal path does not contain subpaths σ satisfying (2), the restriction of e gij to any legal path is an embedding, and hence an isometry, and it follows that fg restricts to an isometry on all legal paths. ♦
Next we review the well known fact that there is a kind of action on the R-tree T+φ by the set of automorphisms of Fr representing a given fully irreducible outer automorphism φ. This set is not a group, instead it is a coset of the normal subgroup Inn(Fr ) in the group Aut(Fr ) — formally this coset is identified with the outer automorphism φ ∈ Out(Fr ) = Aut(Fr )/ Inn(Fr ). So while it does not make sense to ask for a “group action” of the coset, it does make sense to ask for an “Inn(Fr )-affine action”, which is what item (2) gives us:
Theorem 2.15. Let φ ∈ Out(Fr ) be fully irreducible, and denote T+ = T+φ . To each Φ ∈ Aut(Fr ) representing φ there is associated a homothety Φ+ : T+ → T+ with stretch factor λ satisfying the following properties: e→G e be the (1) For each irreducible train track representative g : G → G of φ, letting g˘ : G lift corresponding to Φ, we have Φ+ ◦ fg = fg ◦ g˘ (2) For each Φ, Φ′ representing φ, if γ ∈ Fr is the unique element such that Φ′ = iγ ◦ Φ then Φ′+ = tγ ◦ Φ+ . Proof. Choose any train track representative g : G → G of φ, and refer to Figure 1 for e→G e be the lifts associated to Φ, Φ′ , respectively. Refer notations associated to g. Let g˘, g˘′ : G to Figure 2 for notations associated to g˘. Refer again to Figure 2 for notations associated to g˘′ , but with the prime symbol ′ added to these notations. In particular we have identifications e=G e 0 , g˘ = g˘0 , g˘′ = g˘′ . G 0 Let T+∗ be T+ with the metric scaled down by the factor λ and let h : T+ → T+∗ be the ‘identity’ map; thus h contracts length uniformly by λ. By the universal property of direct ei → T ∗ there is a map Φ∗ : T+ → T ∗ such that limits applied to the maps h ◦ fgi ◦ g˘i : G + + + ∗ h ◦ fgi ◦ g˘i = Φ+ ◦ fgi . Let Φ+ : T+ → T+ be h−1 ◦ Φ∗+ . It follows that fgi ◦ g˘i = Φ+ ◦ fgi , which implies (1) by taking i = 0. Similarly Φ′+ is defined satisfying fgi ◦ g˘i′ = Φ′+ ◦ fgi , which implies (2) together with the fact that g˘i′ = tγ ◦ g˘i and with the universal property of direct limits. e0 such that f0 (˜ Given z1 and z2 in T+ , choose x ˜j ∈ G xj ) = zj . By the third item of e Lemma 2.7, the projection into Gi of the path in Gi connecting e gi0 (x1 ) to gei0 (x2 ) splits 31
into subpaths that are either legal or indivisible Nielsen paths for all sufficiently large i. Corollary 2.14 implies that the distance between z1 and z2 is the sum of the lengths of the legal subpaths of Gi in this splitting. Since g expands length by a factor of λ on all legal subpaths, (1) implies that Φ+ must expand the distance between z1 and z2 by this same factor and since z1 and z2 were arbitrary, it follows that Φ+ expands length uniformly by λ. This proves that Φ+ is a homothety onto its image with stretch factor λ, but Φ+ is surjective by minimality of the Fr -tree T+ . ♦ Next we describe the connection between principal automorphisms representing a fully irreducible φ ∈ Out(Fr ) and the set of branch points of T+φ . Lemma 2.16. Suppose that φ ∈ Out(Fr ) is fully irreducible and rotationless, g : G → G is e→G e is a train track representative, Φ is a principal automorphism representing φ and g˘ : G the principle lift corresponding to Φ, and Φ+ : T+ → T+ is as in Theorem 2.15. Then (1) fg (Fix(˘ g )) is a branch point b of T+ , and Fix(Φ+ ) = {b}. (2) Every direction based at b is fixed by DΦ+ and has the form Dfg (d) where d is a fixed direction based at some v˜ ∈ Fix(˘ g ). (3) If fg (˜ x) = b and d is a direction based at x ˜ then • For some m > 0, g˘m (˜ x) ∈ Fix(˘ g ). • fg (˘ g k (˜ x)) and Dfg (D˘ g k (d)) are independent of k ≥ 0. (4) The assignment Φ 7→ b = Fix(Φ+ ) of (1) defines a bijection between PA(φ) and the set of branch points of T+ . Proof. First we prove (1). Lemma 2.11 implies that Fix(˘ g ) 6= ∅ and that the elements of Fix(˘ g ) are principal vertices. By Corollary 2.14 and Theorem 2.15, the point b = fg (˜ v) is independent of v˜ ∈ Fix(˘ g) and {b} = Fix(Φ+ ), and there is an induced map Dfg from directions at v˜ to directions at b that is injective on the set of fixed directions. If there are at least three fixed directions at v˜ then b is a branch point. Otherwise, there is an indivisible Nielsen path σ and a lift σ ˜ with one endpoint at v˜ and the other at some principal w ˜ ∈ Fix(˘ g ). Let d1 and d2 be distinct fixed directions at v˜ with d2 being the initial direction of σ ˜ and let d3 and d4 be distinct fixed directions at w ˜ with d3 being the initial direction of the inverse of σ ˜ . Then Dfg (d1 ) 6= Dfg (d2 ) = Dfg (d3 ) 6= Dfg (d4 ). To prove that b is a branch point, it suffices to prove that Dfg (d1 ) 6= Dfg (d4 ). If this fails then there exist x ˜ close to v˜ in the direction d1 and y˜ close to w ˜ in the direction d4 such that fg (˜ x) = fg (˜ y ). The second item of Lemma 2.7 implies that the projection into G of the path connecting x ˜ to y˜ splits as a non-trivial legal path followed by the indivisible Nielsen path σ followed by a non-trivial legal path. Corollary 2.14 then implies that x ˜ and y˜ have distinct f˜g -images and this contradiction completes the proof of item (1). We now turn to (3). Choose v˜ ∈ Fix(˘ g ). Corollary 2.14 implies that if fg (˜ x) = b then there exists m ≥ 0 such that either g˘m (˜ x) = v˜ or the path connecting g˘m (˜ x) to v˜ is the lift of a Nielsen path. In either case, g˘m (˜ x) ∈ Fix(˘ g ), which implies that fg (˘ g k (˜ x)) = b for all m k ≥ 0. After increasing m, we may assume that D˘ g (d) ∈ Fix(D˘ g ) for all directions d based at x˜. From the fact that fg ◦ g˘ = Φ+ ◦ fg we conclude that DΦk+ ◦ Dfg (d) = Dfg (D˘ g k (d)) is 32
independent of k ≥ m. Since Φ+ is a homeomorphism, they are in fact independent of k ≥ 0. This completes the proof of item (3) and also implies that Dfg (d) ∈ Fix(DΦ+ ). Since every direction at b occurs as Dfg (d) for some x ˜ and d, we have also proved item (2). Next we prove injectivity of the map Φ 7→ b. Suppose that v˜1 and v˜2 are distinct principal vertices and that fg (˜ v1 ) = fg (˜ v2 ). Let g˘1 be the lift that fixes v˜1 . Since principal vertices of e e is a bijection. G project into Fix(g), the restriction of g˘1 to the set of principal vertices of G k k It follows that g˘ (˜ v1 ) 6= g˘ (˜ v2 ) for all k. The path σ ˜ connecting v˜1 to v˜2 cannot project to a pre-Nielsen path σ which is not Nielsen. Indeed if it did, then for some k > 0, the Nielsen k k k path σ ′ = f# (σ) has the same endpoints as σ and f# (σ ′ ) = f# (σ) which contradicts that σ ′ 6= σ and the fact that f is a homotopy equivalence. Corollary 2.14 therefore implies that σ ˜ projects to a Nielsen path and hence that v˜2 ∈ Fix(˘ g1 ). This proves that Φ 7→ b is injective. To see that Φ 7→ b is onto it suffices to show that if b is a branch point then fg (˜ c) = b for some principal vertex c˜; one then chooses Φ to be the automorphism corresponding to e and directions dl at a the lift of g that fixes c˜. For 1 ≤ l ≤ 3 there are points a ˜l ∈ G ˜l such that fg (˜ al ) = b, and such that Dfg (dl ) are three distinct directions at b. Let aj ∈ G be the projected image of a ˜j . On the one hand, if g m (a1 ) = g m (a2 ) = g m (a3 ) for some m ≥ 0 then there are three directions at this point that have distinct Dfg -images and so must have distinct Dg k -images for all k ≥ 0. Otherwise, Corollary 2.14 implies that g m (aj ) is the endpoint of a Nielsen path for some j and some m ≥ 0. In either case c := g m (aj ) is m a principal, and hence fixed, vertex. There is a path σ connecting aj to c so that g# (σ) is trivial. The lift σ ˜ with endpoint a ˜j terminates at a principal vertex c˜ and g˘(˜ c) = g˘(˜ aj ) = b. This completes the proof of (4) and so the proof of the lemma. ♦
2.6
Geodesic laminations in trees and marked graphs
Given a topological space S let D2 S denote the double space of S, the space of distinct ordered pairs in S: D2 S = {(ξ, η) ∈ S × S ξ 6= η}
The group Z/2 acts freely on D2 S by permuting coordinates. Define the geodesic leaf space of Fr , denoted GFr , to be the set of distinct unordered pairs of points in ∂Fr , that is, the quotient space of the action of Z/2 on D2 ∂Fr : GFr = (D2 ∂Fr ) Z/2
The space GFr is locally a Cantor set. We will usually abuse ordered pair notation and write (ξ, η) for a point in GFr . The action of Fr on ∂Fr induces an action on GFr which has a dense orbit. The quotient space GFr /Fr is therefore a non-Hausdorff topological space; this space is denoted B in [BFH00]. Given a free, simplicial Fr -tree T , we also define the geodesic leaf space 2 GT = (D ∂T ) (Z/2) The natural equivariant homeomorphism ∂Fr ≈ ∂T induces a natural equivariant homeomorphism GFr ≈ GT , and we will henceforth identify GT with GFr . 33
It will be convenient to think of GT not just in the abstract as a set of unordered distinct pairs of boundary points, but more concretely as a set whose elements are bi-infinite geodesics, and even more concretely as a lamination in the sense of topology or dynamics, which is a topological space decomposed into 1-manifolds fitting together locally as an interval crossed with a transversal. This approach is used, for example, in [CP94]. One advantage of this approach is that it avoids conceptual difficulties of non-Hausdorff spaces. We briefly review this point of view. Define a parameterized geodesic in T to be an isometric embedding γ : R → T , and define the geodesic flow of T , denoted GF T , to be the space of parameterized geodesics with the compact open topology. The group R acts homeomorphically on GF T by translating the domain: γ · r(t) = γ(t + r). An orbit of the R action on GF T is called a geodesic flow line, and there are natural bijections between the set of geodesic flow lines, the set of oriented bi-infinite geodesics, and the set D2 ∂T . The geodesic flow lines decompose GF T , and they fit together locally like the product of an interval crossed with a Cantor set, that is, a 1-dimensional lamination with Cantor cross sections. The group Z/2 acts on GF T by reflecting the domain R through the origin. This action is free on the set of geodesic flow lines, interchanging a flow line with its orientation reversal. The quotient GL T = GF T /(Z/2) is called the geodesic lamination of T . An element of GL T can be naturally identified with a “pointed geodesic” in T , a bi-infinite geodesic equipped with a base point. The decomposition of GF T into flow lines descends to a decomposition of GL T into sets called geodesic leaves, where each leaf corresponds to different choices of base points along some bi-infinite geodesic in T ; the geodesic leaves form a 1-dimensional lamination with Cantor cross sections. We obtain natural bijections between the set of geodesic leaves, the set of unoriented bi-infinite geodesics, and the set GT . The action of Fr on GT is free, properly discontinuous, and cocompact, and the quotient space GT /Fr inherits the structure of a 1-dimensional lamination with Cantor cross sections. We can view this more intrinsically as the geodesic lamination of the quotient marked graph G = T /Fr , as follows. We can define the geodesic flow GF G as above, as parameterized local geodesics γ : R → G with the compact open topology, equipped with the R action by translation of the domain. The group Z/2 acts by reflecting the domain through the origin, and we obtain the quotient GL G = GF G/(Z/2). Both GF G and GL G are 1-dimensional laminations with Cantor cross sections, and there are natural identifications GF G ≈ GF T /Fr , GL G = GL T /Fr . Given λ = (ξ, η) ∈ GFr let λT denote the leaf of GL T with ideal endpoints ξ, η; we say that λT is the realization of λ ∈ GFr . This relation is an equivariant bijection between points of GFr and leaves of GL T . Moreover this relation respects the topologies in the following sense: a sequence λi ∈ GFr converges to λ ∈ GFr if and only if the sequence (λi )T converges to λT in the Gromov-Hausdorff topology on closed subsets of T , meaning that the intersections with any closed ball in T converge in the Hausdorff topology on closed subsets of that ball. A sublamination of GFr is a closed, Fr -equivariant subset of GFr . Given a free, simplicial Fr -tree T , a sublamination of GL T is a closed, Fr -equivariant union of leaves of GL T . And, given a marked graph G, a sublamination of GL G is a closed union of leaves of GL G. For any T and for G = T /Fr , and for any sublamination Λ ⊂ GFr , the realization of Λ in T is
34
the sublamination of GL T defined by ΛT = {λT λ ∈ Λ}
Also, the realization of Λ in G is the quotient ΛG = ΛT /Fr , a sublamination of GL G. This gives bijections between sublaminations of GFr , sublaminations of GL T , and sublaminations of GL G. We make free use of these bijections without comment, and we freely use the symbol Λ for sublaminations in any of these contexts. When Λ stands for a sublamination of GFr then we sometimes use Λ(T ) or Λ(G) instead of ΛT or ΛG . When Λ stands for a sublamination e of GL G, the corresponding sublamination of GL T is usually denoted Λ. A sublamination Λ ⊂ GFr is minimal if and only if every Fr -orbit in Λ is dense in Λ; equivalently, for each free simplicial Fr -tree T , each orbit of leaves of ΛT is dense in ΛT ; equivalently, for each marked graph G, each leaf of ΛG is dense in ΛG . The group Aut(Fr ) acts on ∂Fr and therefore also on the geodesic leaf space GFr . Note that this action agrees, under restriction, with the given action of Fr ≈ Inn(Fr ) on GFr . Choosing a marked graph G with universal cover T , GFr is identified with the set of leaves of GL T and so Aut(Fr ) acts on the latter, in agreement under restriction with the action of Fr ≈ Inn(Fr ). Modding out by Fr ≈ Inn(Fr ) we obtain an action of Out(Fr ) on the set of leaves of GL G. We will use these actions without reference in what follows.
2.7
The expanding lamination Λ−
The chief example of a sublamination of GFr is the expanding lamination Λ− of a fully irreducible outer automorphism φ ∈ Out(Fr ), as constructed in [BFH97]. We review here the construction, well-definedness, and properties of Λ− . In particular we present key features of Λ− in Lemmas 2.17 and 2.18, the second of which motivates our choice of notation. The construction of Λ− will be revisited in Section 7.4 when we extend the construction to a more general context. Complete details on Λ− can be found in either section 1 of [BFH97] or section 3 of [BFH00]. Definition and basic properties of Λ− . Consider any train track representative g : G → G of φ. Choose a sequence of pairs (Ei , xi ) so that each Ei is an edge of G, xi ∈ int(Ei ), and we assume that λi d(xi , ∂Ei ) → +∞; this is true, for example, if (xi ) is a periodic orbit. Then (g i (Ei ), g i (xi )) is a sequence of pointed geodesic arcs whose lengths go to +∞ so that the distance from the base point to the endpoints also goes to +∞. We may therefore choose a subsequence of (Ei , xi ) so that (g i (Ei ), g i (xi )) converges to a bi-infinite, pointed geodesic in G. The union of such geodesics, over all the choices made, is clearly a sublamination of GL G that we denote Λ− (g). In [BFH97] it is proved that the expanding lamination Λ− ⊂ GFr of φ is well-defined and that its realization in G, for each train track representative g : G → G of φ, is Λ− (g). If the train track map is clear from the context then we will simply denote Λ− (g) as Λ− (φ) or just as Λ− . The expanding lamination for φp , p > 0, is evidently the same as the expanding lamination for φ so we may replace φ by an iterate when it is convenient. By irreducibility of g : G → G, every leaf or half-leaf of Λ− (g) covers every edge of G. Also, Λ− is carried by no proper free factor of Fr (see Example 2.5 (1) in [BFH97]), so the realization of Λ− in any marked graph has the property that every leaf covers every edge. 35
Action of φ on Λ− . Under the action of Out(Fr ) on the set of leaves of GL G described at the end of Section 2.6, the outer automorphism φ leaves invariant the set of leaves of Λ− . We denote this restricted action by ℓ 7→ φ# (ℓ), for each leaf ℓ of Λ− . This action is characterized by saying that for each train track representative g : G → G of φ and each leaf ℓ of Λ− , g takes the realization of ℓ in G to the realization of φ# (ℓ) in G. For any Φ ∈ Aut(Fr ) that ˜ which is e − , denoted ℓ˜ 7→ Φ# (ℓ), represents φ, there is an action of Φ on the set of leaves of Λ e→G e well-defined by saying that for each train track representative g : G → G, letting g˘ : G ˜ e be the lift of g corresponding to Φ, the map g˘ takes the realization of ℓ in G to the realization ˜ in G. e To justify the existence of these actions see the final comments of Section 2.6. of Φ# (ℓ) e − is periodic if A leaf ℓ of Λ− is periodic if φi# (ℓ) = ℓ for some i ≥ 1, and a leaf ℓ˜ of Λ ˜ ˜ Periodicity of there exists Φ ∈ Aut(Fr ) representing an iterate of φ such that Φ# (ℓ) = ℓ. e − to Λ− . leaves is preserved under projection from Λ Λ− and local Whitehead graphs. From the definition of Λ− (g), for any x ∈ G the local Whitehead graph W (x; G) has a description in terms of Λ− (g), as follows. A nondegenerate turn (d1 , d2 ) at x is taken by some leaf of Λ− (g) if and only if this turn is taken by g k (E) for some edge E of G and some k ≥ 1, which holds if and only if (d1 , d2 ) are the endpoints of an edge of the local Whitehead graph W (x; G).
Singular leaves of Λ− and stable Whitehead graphs. Next we introduce notation and terminology to refer to certain leaves of Λ− , called “singular leaves” in analogy to singular leaves of the unstable foliation of a pseudo-Anosov surface homeomorphism. These singular leaves can be used to relate Λ− to stable Whitehead graphs. Let g : G → G be an irreducible train track map representing an iterate of φ such each principal or nonprincipal periodic vertex v is fixed and each periodic direction at v is fixed. e and a lift g˘ : G e→G e that fixes v˜. If d is a direction at v˜ fixed by D˘ Choose a lift v˜ ∈ G g and e is the edge at v˜ corresponding to d, then g˘j (E) e is a proper initial subpath of g˘j+1 (E) e for E e is a ray in G e converging to some P ∈ FixN (ˆ e = ∪∞ g˘j (E) g ). We say each j ≥ 0 and so R j=0 e e e that R is the ray determined by d (or by E), and if v is principal then R is a singular ray. e1 , R e2 be rays at v˜ determined by distinct fixed directions d1 , d2 , corresponding to Let R e1 , E e2 at v˜. We claim that if d1 and d2 are endpoints of an edge in SW (˜ e (which edges E v ; G) ˜ e e e e This is always true if v˜ is not principal), then ℓ = R1 ∪ R2 is a leaf of Λ− realized in G. l e e e follows because there exists an edge E of G and l > 0 such that g˘ (E) contains the turn e converge (in the Gromov–Hausdorff topology (d1 , d2 ), and so as i → +∞ the paths g˘l+i (E) ˜ e e on closed subsets) to ℓ = R1 ∪ R2 . The leaf ℓ˜ is called the leaf at v˜ determined by d1 and d2 e1 and E e2 ). If v˜ is principal then we say that ℓ˜ is a singular leaf. (or by E e1 ∪ R e2 is a In the previous paragraph, the converse to the claim holds as well: if ℓ˜ = R e − realized in G e then d1 , d2 bound an edge of SW (˜ e This holds because, by leaf of Λ v ; G). e − , each segment of ℓ˜ must be contained in g˘l (E) e for some edge E e and some definition of Λ e l > 0. It follows that SW (˜ v ; G) can be identified with the graph having one vertex for each e based at v˜ and having an edge connecting the vertices corresponding to a pair singular ray R e1 , R e2 at v˜ if and only if R e1 ∪ R e2 is a singular leaf at v˜. of singular rays R 36
Characterizing Λ− . The next lemma gives a useful characterization of Λ− , and its corollary characterizes the nonsingular φ-periodic leaves of Λ− . Lemma 2.17. Λ− is the unique minimal φ-invariant lamination whose leaves are legal with respect to every train track representative g : G → G of φ. Every leaf ℓ˜ constructed as above, and in particular every singular leaf, is contained in Λ− . Proof. With notation as above, consider the set L of minimal φ-invariant laminations whose leaves are legal with respect to every train track representative g : G → G of φ. That Λ− ∈ L follows from the construction of Λ− in section 1 of [BFH97], as reviewed above. We will show that if Λ′ ∈ L and ℓ˜ is any leaf constructed as above, then ℓ˜ is a leaf of Λ′ . Since there can only be one minimal lamination that contains ℓ˜ it follows Λ− is the only element of L. e→G e as above, and choose v˜ ∈ G e Pass to a power and choose an appropriate lift g˘ : G e1 and E e2 at v˜ all of which are fixed by g˘, so that E e1 , E e2 determine an and distinct edges E e There is an edge E e and l > 0 such that g˘l (E) e contains the turn (E e1 , E e2 ). edge in SW (˜ v ; G). ′ ′ ′ ˜ e e Since g is irreducible and leaves of Λ are legal in G, there is a leaf ℓ of Λ that contains E. −1 e −1 e 2 e −1 ˜ e e Each path in the nested increasing sequence E1 E2 ⊂ g˘(E1 E2 ) ⊂ g˜ (E1 E2 ) ⊂ · · · occurs e ′ . It follows that their union ℓ˜ is a leaf of Λ e ′. as a subpath of a leaf of Λ ♦ ˆ = FixN (Φ) ˆ = Corollary 2.18. If Φ ∈ Aut(Fn ) represents an iterate of φ and PerN (Φ) e Conversely, {P1 , P2 } then the line ℓ˜ bounded by P1 and P2 is a periodic nonsingular leaf of Λ. ˜ e if ℓ is a nonsingular periodic leaf of Λ then there exists Φ ∈ Aut(Fn ) representing an iterate ˜ ˆ = FixN (Φ) ˆ = ∂ ℓ. of φ such that PerN (Φ)
Proof. Choose an irreducible train track map g : G → G representing φ and let g˘k be the lift of an iterate of g that corresponds to Φ. Since P1 and P2 are attractors, Fix(˘ gk ) 6= ∅; see for example Lemma 3.9 and Lemma 3.13 of [FH04]. After replacing g˘k by an iterate we may choose x ˜ ∈ Fix(˘ gk ) and a pair of fixed directions d1 and d2 at x ˜. After subdividing G, we may assume that x ˜ is a vertex. If there were any other periodic direction at x ˜ then ˆ in addition to P1 , P2 , the corresponding singular ray would give another point in PerN (Φ) e has just the two a contradiction. It follows that the stable Whitehead graph SW (˜ x; G) vertices d1 , d2 , and moreover the rays R1 , R2 determined by d1 , d2 must end at points of ˆ and up to re-indexing these points must be P1 , P2 . Since SW (˜ e is connected PerN (Φ), x; G) e (see Section 2.4), there must be an edge of SW (˜ x; G) connecting d1 to d2 . An application e of Lemma 2.17 shows that ℓ˜ = R1 ∪ R2 is in Λ. e Choose Φ ∈ Aut(Fr ) repreFor the converse, let ℓ˜ be a nonsingular periodic leaf of Λ. ˜ = ℓ. ˜ Choose an irreducible train track map g : G → G senting an iterate of φ so that Φ#(ℓ) e→G e be the lift of an iterate of g that corresponds to Φ. Using representing φ and let g˘k : G ˜ ˜ = ℓ. ˜ Since g˘k expands ℓ˜ by the uniform e we have g˘k (ℓ) ℓ to denote its own realization in G, ˜ factor λ, there is a point x ∈ ℓ fixed by g˘k . Subdivide G so that x is a vertex. Up to passing to the square, D˘ gk fixes the two directions d1 , d2 of ℓ˜ at x, and the rays determined by these ˆ ⊂ PerN (Φ). ˆ two directions are the two rays of ℓ˜ based at x. It follows that ∂ ℓ˜ ⊂ FixN (Φ) ˆ If there were another point in PerN (Φ) then, passing to a power so that this point is in ˆ it would follow that ℓ˜ is singular, contrary to hypothesis. FixN (Φ), ♦
37
2.8
Relating Λ− to T− and to T+
Consider an R-tree T representing a point in ∂Xr , and a nonempty minimal sublamination Λ of GFr . Following [BFH97] section 3 we say that Λ has length zero in T if for every G ∈ Xr e → T there exists C ≥ 0 such that for every leaf and every Fr -equivariant morphism h : G e e ℓ ⊂ G of ΛG we have diamT− (h(ℓ)) ≤ C. Lemma 2.19. For each fully irreducible φ ∈ Out(Fr ), Λ− = Λ− (φ) is the unique minimal lamination that has length zero in T− = T− (φ).
Proof. This is proved in Lemma 3.5 (4) of [BFH97], but restricted to the so-called “irreducible laminations”, meaning the set of expanding laminations of all fully irreducible outer automorphisms in Fr . However, the proof applies to arbitrary minimal sublaminations, for as is noted several times in [BFH97], the only property of irreducibility of Λ that is really used is quasiperiodicity of leaves of Λ, which is equivalent to minimality of Λ. See for example the proof of [BFH97] Proposition 2.12 which is used in the proof of Lemma 3.5 (4). ♦ Remark Lemma 2.19 shows that the expanding lamination of a fully irreducible outer automorphism φ is naturally related to the repelling tree T− rather than the attracting tree T+ . This is why we have chosen to denote the expanding lamination with a subscript “−”, altering the notation from the subscript “+” used in [BFH97]. We next relate Λ− to T+ . e − , the Lemma 2.20. For any train track representative g : G → G of φ and any leaf ℓ˜ of Λ ˜ e e restriction of the map fg : G → T+ to the realization of ℓ in G is an isometry. Moreover, the bi-infinite line in T+ which is the image of this isometry is independent of the choice of ˜ this image is called the realization of ℓ˜ in T+ . g, depending only on ℓ; ♦ Motivated by the first part of this lemma, given a free, simplicial Fr -tree T , an Fr equivariant map f : T → T+ is said to be a Λ− isometry, or to be Λ− isometric, if the e − realized in T is an isometry. The first part of the lemma restriction of f to any leaf of Λ therefore says that the map fg is a Λ− isometry. Remark. The property that T has a Λ− isometry f : T → T+ does not, in and of itself, determine the map f . In Section 5.4 we give an example of a marked graph G such that e → T+ — in fact, that example will have a there exist at least two distinct Λ− isometries G one-parameter family of distinct such maps. On the other hand, the arguments in Section 5.4 e goes in T+ . will show that a Λ− isometry is determined by where a single point of G
Proof of Lemma 2.20. The first part follows by combining Corollary 2.14 and Lemma 2.17. e − , we must prove that To prove uniqueness of the realization in T+ of a leaf ℓ˜ of Λ ˜ ˜ fg (ℓ) is well-defined independent of g. Think of ℓ abstractly as an element of GFr , and ˜ For any train track choose a sequence γi ∈ Fr whose abstract axes A(γi ) in GFr limit to ℓ. representative g : G → G of φ, the realizations A(γi )Ge converge to the realization ℓ˜G in the e It follows that the sets pi = ℓ˜G ∩ A(γi ) e Gromov–Hausdorff topology on closed subsets of G. G form a longer and longer family of paths that exhausts ℓ˜G . While A(γi )Ge need not be legal 38
e and so may not embed in T+ under fg , its subpath pi is legal and does embed in T+ . in G By bounded cancellation, Lemma 2.5, there exists B independent of i and depending only on fg , such that if p′i is obtained from pi by cutting off segments of length B at both ends, then fg (p′i ) embeds in A(γi )T+ , the axis of γi in T+ . But the sequence of paths fg (p′i ) form a longer and longer family of paths in T+ that exhausts fg (ℓ˜G ). It follows that A(γi )T+ approaches fg (ℓ˜G ) in the Gromov-Hausdorff topology on closed subsets of T+ . Since the ˜ in T+ follows. ♦ sequence γi is independent of the choice of g : G → G, uniqueness of fg (ℓ) We conclude with a relation between Λ− and the homothety Φ+ : T+ → T+ . e − , the map Φ+ Corollary 2.21. For each Φ ∈ Aut(Fn ) representing φ, and each leaf ℓ˜ of Λ ˜ in T+ . takes the realization of ℓ˜ in T+ to the realization of Φ# (ℓ) e→G e the lift corresponding Proof. Let g : G → G be a train track representative of φ and g˘ : G ˜ ˜ e to Φ. Let ℓ denote its own realization in G, and so fg (ℓ) is its realization in T+ . Also, the ˜ in G ˜ and its realization in T+ is fg (˘ ˜ But we know that e is g˘(ℓ), realization of Φ# (ℓ) g (ℓ)). fg ◦ g˘ = Φ+ ◦ fg
˜ in T+ . and so Φ+ takes the realization of ℓ˜ in T+ to the realization of Φ# (ℓ) ♦
39
3
The ideal Whitehead graph
On a closed oriented surface S, the topological structure of the unstable foliation of a pseudoAnosov homeomorphism φ : S → S near a singular point s ∈ S is fully determined by the number n ≥ 3 of unstable prongs at s. Associated to s there is a principal region (that is, a complementary component) of the unstable geodesic lamination of φ, this region is an ideal n-gon, and its n sides are the singular leaves associated to s, which are circularly ordered by the asymptote relation. Also, in the R-tree dual to the unstable foliation, associated to each n-pronged singularity s there is an orbit of branch points of the R-tree, and the number of directions at each such branch point is exactly n. The singularity structure of φ is therefore described fully by a list of integers ni giving the number of unstable prongs at each singularity. The singularity structure is constrained by the Euler-Poincar´e index P equation (ni − 2) = −2χ(S), and the paper [MS93] proves that outside of a few fully understood exceptions that occur only for a few surfaces of low topological complexity, the index equation gives a complete description of all singularity structures that actually occur for pseudo-Anosov homeomorphisms of S. Singularity structures of nongeometric, fully irreducible outer automorphisms φ ∈ Fr are less well understood. The most that is known so far is an index inequality of Euler-Poincar´e type that is proved in [GJLL98]: there are finitely many Fr orbits of branch points in T+φ , each branch point has finitely manyP directions, and if ni is the number of directions at each branch point in the ith orbit then (ni − 2) ≤ 2(r − 1). Whether this gives a complete description of all singularity structures that actually occur, along the lines of [MS93], has not been investigated. In the index inequality of [GJLL98], the significant information about a singularity is, like in the surface case, a raw count of its number of “unstable directions”, encoded as directions in T+φ . In this section we study a finer invariant of the singularity structure of a nongeometric, fully irreducible φ ∈ Out(Fr ), the “ideal Whitehead graph” of φ, which captures the asymptotic relations among singular leaves of Λ− (φ). The ideal Whitehead graph is first defined in terms of principal automorphisms representing φ or, equivalently, in terms of principal lifts of a train track representative. The main results of this section give alternative descriptions of the ideal Whitehead graph: in terms of singular leaves of train track representatives (Lemma 3.1 and Corollary 3.2); next in terms of the asymptote relation among singular leaves (Lemma 3.3); and finally in terms of branch points in the R-tree T+φ (Lemma 3.4). To describe the ideal Whitehead graph very roughly, to each “singularity of φ” there is associated a component of the ideal Whitehead graph, a finite graph with singular leaves as edges, and with vertices describing asymptotic relations between singular leaves. By analogy, for a pseudo-Anosov surface singularity with n unstable prongs, the corresponding component of the ideal Whitehead graph has n vertices and n edges forming a topological circle. For a general nongeometric, fully irreducible φ ∈ Out(Fr ), components of the ideal Whitehead graph need not be topological circles; indeed, they may have cut points, as we show by example in Section 3.4. For present purposes we restrict to the case that φ is rotationless, but the reader will easily see how to proceed without this restriction. Also, we shall restrict attention to an Fr equivariant version of the ideal Whitehead graph; each component having trivial stabilizer, the reader will easily see how to pass to a quotient under Fr to obtain a nonequivariant version. 40
For the rest of this section we fix a nongeometric, fully irreducible, rotationless φ ∈ Out(Fr ), and we denote Λ− = Λ− (φ) and T+ = T+φ . All automorphisms and all train track maps will be representatives of φ.
3.1
Definition and structure of the ideal Whitehead graph
Recall that, abstractly, leaves of Λ− are just points (P1 , P2 ) ∈ D2 ∂Fr . For every principal automorphism Φ let L(Φ) be the set of leaves (P1 , P2 ) of Λ− with ˆ P1 , P2 ∈ FixN (Φ). Define the component W (Φ) of the ideal Whitehead graph determined ˆ and an edge connecting by Φ to be the graph with one vertex for each point in FixN (Φ), ˆ P1 , P2 ∈ FixN (Φ) if there is a leaf (P1 , P2 ) ∈ L(Φ). Corollary 2.9 implies that for each point ˆ and so P ∈ ∂Fr there is at most one principle automorphism Φ such that P ∈ FixN (Φ), L(Φ1 ) ∩ L(Φ2 ) = ∅ for distinct principle automorphisms Φ1 , Φ2 . It therefore makes sense to define the ideal Whitehead graph W (φ) to be the disjoint union of its components W (Φ), one for each principal automorphism Φ representing φ. The definition of W (φ) can be easily reformulated in terms of any train track representae → G, e let L(˘ ˜− tive g : G → G of φ. For every principal lift g˘ : G g ) denote the set of leaves of Λ with both ideal endpoints in FixN (ˆ g ), define W (˘ g ) to be the graph with one vertex for each point in FixN (ˆ g ) and an edge connecting the vertices corresponding to P1 , P2 ∈ FixN (ˆ g ) if there is a leaf ℓ˜ ∈ L(˘ g ) with ideal endpoints P1 and P2 , and define W (g) to be the disjoint union of the W (˘ g ). Note that if Φ is the principal automorphism corresponding to g˘ then there are natural isomorphisms L(˘ g ) ≈ L(Φ) and W (˘ g ) ≈ W (Φ), and there is a natural isomorphism W (g) ≈ W (φ). Formally, it is convenient to let the vertex set of W (˘ g ) be FixN (ˆ g) and to let the edge interiors of W (˘ g ) be the disjoint union of the leaves of L(˘ g). e→G e and v˜ ∈ Fix(˘ For any principal lift g˘ : G g ), we identify the local stable Whitehead graph SWv˜ with a subgraph of W (˘ g ) as follows. Each vertex di of SWv˜ is a direction at v˜ fixed by D˘ g , di is the initial direction of a unique singular ray Ri , and di is identified with the vertex of W (˘ g ) which is the ideal endpoint Pi of the ray Ri . As shown in Section 2.7, two vertices di , dj of SWv˜ that are connected by an edge of SWv˜ correspond to a singular leaf ℓ = R1 ∪ R2 with ideal endpoints P1 , P2 , and so ℓ corresponds to an edge of W (˘ g ) with endpoints P1 , P2 . Lemma 3.1. Let g˘ be a principal lift of g. (1) Every P ∈ FixN (ˆ g ) is the endpoint of at least one element of L(˘ g). (2) Singular rays R1 at v˜1 ∈ Fix(˘ g ) and R2 at v˜2 ∈ Fix(˘ g) terminate at the same point P ∈ Fix(ˆ g ) if and only if the path σ ˜ connecting v˜1 to v˜2 projects to an indivisible Nielsen path σ and determines the same direction at v˜i as does Ri . (3) L(˘ g ) is the disjoint union, over all v˜ ∈ Fix(˘ g ), of the singular leaves at v˜. Proof. Item (1) follows from Lemma 4.7 of [FH04]. The path σ of (2) is a Nielsen path. If R1 and R2 eventually coincide then σ ˜ =α ˜ β˜−1 ˜ where α ˜ ⊂ R1 and β ⊂ R2 . This proves the only if part of (2). For the if part, suppose that σ = αβ¯ is indivisible. Let α ˜ and β˜ be the lifts of α and β that begin at v˜1 and v˜2 ˜ = β˜τ˜n . respectively. For all n ≥ 1, there are paths τ˜n such that g˘n (˜ α) = α ˜ τ˜n and g˘n (β)
41
˜ are an The paths g˘n (˜ α) are an increasing sequence whose union is R1 and the paths g˘n (β) increasing sequence whose union is R2 . This completes the proof of (2). Each singular leaf based at an element of Fix(˘ g ) is contained in L(˘ g ) by construction. Conversely, since the endpoints of any ℓ˜ ∈ L(˘ g) are attractors for gˆ, there is a necessarily unique fixed point x ˜ ∈ ˜l; see for example Lemma 3.9 and Lemma 3.10 of [FH04]. Lemma 2.11 implies that x ˜ is a principal vertex. The turn taken by ℓ˜ at x ˜ determines an edge of SWx˜ and hence a singular leaf ℓ˜′ at x ˜. If ℓ˜ 6= ℓ˜′ then there are rays R and R′ based at x˜ with fixed point free interiors, with a common initial edge and with distinct terminal endpoints P and P ′ in FixN (ˆ g ). The last point of R ∩ R′ is therefore mapped by g˘ both to R − R′ and ′ to R − R, a contradiction. This proves that ℓ is a singular leaf at x˜ and completes the proof of (3). ♦ The following corollary says that each component of the ideal Whitehead graph is a union of local Whitehead graphs, pasted together along cut points: Corollary 3.2. Let g˘ be a principal lift of g. (1) W (˘ g ) is connected. (2) W (˘ g ) = ∪SW (˜ v ) where the union is taken over all v˜ ∈ Fix(˘ g ). (3) For i 6= j, SW (˜ vi ) and SW (˜ vj ) intersect in at most one vertex. If they do intersect at a vertex P , then P is a cut point of W (˘ g ), in fact P separates SW (˜ vi ) from SW (˜ vj ) in W (˘ g ). Proof. (2) follows from the first and third item of Lemma 3.1. (1) follows from (2), item 2 of Lemma 3.1, and the fact that any two elements of Fix(˘ g) are connected by a finite concatenation of indivisible Nielsen paths. To simplify notation, for proving (3) we assume i = 1, j = 2. Suppose that P ∈ FixN (ˆ g) is a common vertex of SW (˜ v1 )∩SW (˜ v2 ). There are singular rays R1 and R2 terminating at P ˜ − contains both v˜1 and v˜2 , it follows and initiating at v˜1 and v˜2 respectively. Since no leaf of Λ that R1 ∩ R2 ∩ {˜ v1 , v˜2 } = ∅. If Q 6= P is another common vertex of SW (˜ v1 ) ∩ SW (˜ v2 ) then there are singular rays R1′ and R2′ terminating at Q and initiating at v˜1 and v˜2 respectively. The initial direction of R1 is distinct from that of R1′ and similarly for R2 and R2′ . It follows e is a tree. This that R1 ∪ R2 ∪ R1′ ∪ R2′ contains an embedded circle, contradicting that G proves SW (˜ v1 ) ∩ SW (˜ v2 ) = {P }. To prepare for proving the last part of (3), let W d (˘ g ) denote W (˘ g ) with vertices deleted, namely, the disjoint union of all singular leaves based at all fixed points of g˘; formally, one can think of W d (˘ g ) as the set of ordered pairs (l, x) where l is a singular leaf based at a fixed e be the natural map, taking each singular leaf to point of g˘ and x ∈ l. Let r : W d (˘ g) → G itself by the identity map; formally, r(l, x) = x. The set r−1 (Fix(˘ g )) consists of exactly one point in each edge of W (˘ g ), called the basepoint of that edge, separating the edge into two open half-edges. Suppose that P does not separate SW (˜ v1 ) from SW (˜ v2 ) in W (˘ g ). Then there are edge basepoints w1 , w2 ∈ W d (˘ g ) with r(wi ) = vi , and there is an embedded path γ in W (˘ g) from w1 to w2 disjoint from P . Let γ d = γ ∩ W d (˘ g ), that is, γ with vertices of W (˘ g) d e deleted. The image r(γ ) is a connected subset of G, because near any deleted vertex Q of γ e asymptotic to Q. Also, since the two ends of γ incident to Q are mapped by r to rays in G 42
γ does not pass through P , and since γ is embedded, it follows that γ does not intersect any open half-edge incident to P , from which it follows that the image r(γ d ) is disjoint from e − (int(R1 ) ∪ int(R2 )) int(R1 ) ∪ int(R2 ). Therefore, r(γ d ) contains an embedded path in G from v˜1 to v˜2 . It follows that r(γ d ) ∪ R1 ∪ R2 contains an embedded circle, contradicting e is a tree. that G ♦
3.2
Asymptotic leaves and the ideal Whitehead graph
˜ − are asymptotic if they have a common infinite end. This genA pair of distinct leaves of Λ ˜ − whose non-trivial classes are called asymptotic erates an equivalence relation on leaves of Λ ˜ −. equivalence classes of Λ Our next lemma relates a component of the ideal Whitehead graph to asymptotic equivalence classes. Lemma 3.3. The map g˘ 7→ L(˘ g ) defines a bijection between the set of principal lifts of g ˜ − . Two ends of edges of W (˘ and the set of asymptotic equivalence classes of Λ g) are incident to the same vertex if and only if the corresponding rays in leaves of L(˘ g) are asymptotic. Proof. The second sentence is obvious. Injectivity of the map g˘ 7→ L(˘ g) follows from the fact that L(˘ g1 ) ∩ L(˘ g2 ) = ∅ for distinct g˘1 and g˘2 . Singular leaves at v˜ ∈ Fix(˘ g ) belong to the same asymptotic equivalence class because SWv˜ is connected. Since any two elements of Fix(˘ g ) are connected by a finite concatenation of indivisible Nielsen paths, the second item of Lemma 3.1 implies that all elements of L(˘ g) are in the same asymptotic equivalence class. e then they are elements It remains to show that if ℓ˜′ and ℓ˜′′ are asymptotic leaves in G e so consider asymptotic leaves of some L(˘ g ). It is convenient to work in G rather than G ℓ′ , ℓ′′ ∈ Λ− . For each j ≥ 0, there are leaves ℓ′j and ℓ′′j such that g j (ℓ′j ) = ℓ′ and g j (ℓ′′j ) = ℓ′′ . The intersection of ℓ′j and ℓ′′j is a ray Rj based at a vertex xj . Denote the rays in ℓ′j and ℓ′′j that are complementary to Rj by Rj′ and Rj′′ respectively. Let Ej , Ej′ and Ej′′ be the initial edges of Rj , Rj′ and Rj′′ respectively and let dj , d′j and d′′j be the directions that they determine. As a first case suppose that (d′j , d′′j ) is legal for all j ≥ 0. Then xj is principal and g i (xj+i ) = xj for all i, j ≥ 0. It follows that xj ∈ Fix(g) is independent of j and we now denote xj by v. The analogous argument shows that Ej , Ej′ and Ej′′ are independent of j and we denote these by E, E ′ and E ′′ . For all j, the paths g j (E) and g j (E ′ ) are respectively initial subpaths of R0 and R0′ . It follows that ℓ′ is the singular leaf at v determined by E and E ′ . The symmetric argument shows that ℓ′′ is the the singular leaf at v determined by E and E ′′ . The point v˜ at which ℓ˜′ and ℓ˜′′ diverge projects to v. If g˘ is the lift that fixes v˜, then ℓ′ and ℓ′′ are singular leaves at v˜ for g˘. This completes the proof in this first case. Assume now that (d′j , d′′j ) is illegal for some, and hence every sufficiently large, j. Item 5 of Lemma 2.7 and Lemma 2.12 imply that there is an indivisible Nielsen path αj β¯j such i that α ¯ j is an initial path of Rj′ , β¯j is an initial path of Rj′′ , and g# (αi+j β¯i+j ) = αj β¯j for all i, j ≥ 0. It follows that αj , βj , xj and the initial endpoints yj and zj of αj and βj respectively are independent of j; denote these by α, β, v, y and z. Arguing as in the previous case we see that ℓ′ and ℓ′′ are singular leaves at y and z respectively. The point v˜ at which ℓ˜′ and ℓ˜′′ diverge projects to the common terminal endpoint v of α and β. Let α ˜ and β˜ be the lifts 43
˜ that terminate at v˜ and let g˘ be the lift that fixes the initial endpoints y˜ and z˜ of α ˜ and β. Then ℓ′ and ℓ′′ are singular leaves for g˘ at y˜ and z˜ respectively. ♦
3.3
T+ and the ideal Whitehead graph
Next we relate the ideal Whitehead graph to T+ . Recall Corollary 2.20 which explains how ˜ − are realized in T+ . the leaves of Λ Suppose that b ∈ T+ is a branch point, Φ ∈ PA(φ) is the principal automorphism e→G e is the corresponding principal lift. Each corresponding to b by Lemma 2.16, and g˘ : G singular leaf ℓ in L(˘ g) passes through some x ∈ Fix(˘ g ), and so fg (ℓ), the realization of ℓ in T+ , passes through the point fg (x) = b; we call these the singular leaves at b. Each singular leaf at b is divided by b into singular rays at b which are the images under fg of the singular rays at some x ∈ Fix(˘ g ). Define the local Whitehead graph at a point z ∈ T+ , denoted W (z; T+ ), to be the graph with one vertex for each direction di at z and with an edge connecting di to dj if the turn e the map Dx fg (di , dj ) is taken by the realization of some leaf of Λ− in T+ . For each x ∈ G e maps the directions of G at x to the directions of T+ at f (b), and this map extends to a e 7→ W (b; T+ ), still denoted Dx fg (though we often drop x from the simplicial map W (x; G) notation Dx fg . With b, Φ, g˘ as above, for each x ∈ Fix(˘ g) the map Dx fg is injective on the set of e ֒→ fixed directions at x, and therefore Dx fg restricts to a simplicial embedding SW (x; G) W (b; T+ ). Lemma 3.4. With b, Φ, and g˘ as above, as x varies over Fix(˘ g ) the embeddings e ֒→ W (b; T+ ) Dx fg : SW (x; G)
induce an isomorphism of W (Φ) with W (b; T+ ). More precisely, (1) Every direction at b is realized by a singular ray. e with basepoints in Fix(˘ (2) Singular rays R1 and R2 in G g ) map to the same singular ray in T+ if and only if they terminate at the same point in FixN (ˆ g ). ˜ − if and only if it is taken by a singular leaf at b. (3) A turn at b is taken by some leaf of Λ e → W (b; T+ ) Proof. Observe first that, by Corollary 3.2 and (2), the maps Dx fg : SW (x; G) for x ∈ Fix(˘ g ) fit together to define a simplicial map W (Φ) 7→ W (b; T+ ), which by items (1)–(3) is an isomorphism. Item (1) follows from the second item of Lemma 2.16 and the fact that every fixed direction at an element of Fix(˘ g ) is realized by a singular ray. The “if” part of (2) follows from the second item of Lemma 3.1 and Corollary 2.14. For the “only if” part, let R1 and R2 be singular rays at v˜1 and v˜2 such that fg (R1 ) = fg (R2 ). We prove that R1 and R2 converge to the same point in Fix(ˆ g). Let di be the direction determined by Ri at v˜i . If v˜1 = v˜2 , then R1 = R2 and the result is obvious. Suppose then that v˜1 6= v˜2 . The path connecting v˜1 to v˜2 splits as a concatenation µ ˜=µ ˜1 · . . . · µ ˜l where each µi is an indivisible Nielsen path. For i = 1, 2 choose x ˜i close to and in the direction di from v˜i such that fg (˜ x1 ) = fg (˜ x2 ). The path µ ˜′ connecting x ˜1 to x˜2 is obtained from µ ˜ by 44
making alterations at the beginning and end. If d1 is the direction determined by µ ˜1 , then the path from v˜1 to x ˜1 is removed from µ ˜1 ; otherwise the path from x ˜1 to v˜1 is added. Similar changes are made at the terminal end. Since there are only finitely many indivisible Nielsen or pre-Nielsen paths in G we may assume that x ˜1 and x ˜2 have been choosen so that µ ˜′ does k k not project to a Nielsen or pre-Nielsen path. Corollary 2.14 implies that g˘ (˜ x1 ) = g˘ (˜ x2 ) for some k > 0. The second item of Lemma 2.7 therefore implies that x ˜1 and x ˜2 belong to the interior of the same µ ˜j . It follows that µ = µ1 is indivisible and that d1 and d2 point into µ ˜. The second item of Lemma 3.1 implies that R1 and R2 terminate at the same endpoint as desired. + ˜ ˜ The “if” part of (3) is obvious. Suppose the turn (d+ 1 , d2 ) is taken by a leaf ℓ of Λ− . e contains some x The realization of ℓ˜ in G ˜ ∈ fg−1 (b). By Lemma 2.16(3), for some m > 0 m ˜ at g˘m (˜ we have g˘ (˜ x) ∈ Fix(˘ g ) and the two directions d1 , d2 of g˘m (ℓ) x) are fixed by D˘ g, + m so fg (g (˜ x)) = b and Dfg (di ) = di . Also, since the turn (d1 , d2 ) is taken by some leaf, ˜ it follows that this turn is taken by some singular leaf ℓ˜′ based at g˘m (˜ namely g˘m (ℓ), x), and + ′ so fg (ℓ˜ ), the realization of ℓ˜′ in T+ , is a singular leaf based at b taking the turn (d+ 1 , d2 ). ♦
3.4
An example of an ideal Whitehead graph
In this section we give an example of a train track map representing a nongeometric, fully irreducible element φ ∈ Out(F3 ), together with a complete computation of (the quotient by F3 of) the ideal Whitehead graph of φ. Example: The graph G and train track map g : G → G are given by A
E
:qo
F
Cp
g:
D
A 7→ AF D 7→ DF AF DF E E→ 7 DF AF DF E F → 7 DF
One easily shows by the method of Stallings folds that g is a homotopy equivalence. Let φ ∈ Out(F3 ) be represented by g, with respect to any marking of G. We show that φ is fully irreducible with ideal Whitehead graph A
D
F
E
Noting that g(p) = g(q) = q, the induced map Dg on directions is computed to be p
:
qY
:
A
AY
FW
E
F
D
/D Y
EW
The only illegal turn is therefore {E, D}. The words ED and DE not occurring in the definition of g, it follows that g is a train track map (of course, any map of a directed graph which preserves direction is a train track map). 45
We verify that g has no periodic Nielsen paths by proving that for any immersed edge path γ its straightened image g# (γ) is legal. Each illegal turn of γ is at the concatenation point of a subword DE or ED, and each such subword is cancelled by g because g(D) = g(E). It suffices therefore to show that when the DE and ED subwords of g(γ) are cancelled, the turn that results from each such cancellation is legal. Consider one such subword, represented up to reversal by writing γ = γ1 DEγ2 . It suffices to consider the case where γ1 , γ2 are nontrivial and so we can write γ = γ3 XDEY γ4 where X, Y are edges. Noting that g(X) does not end with E nor does g(Y ) begin with E, it follows that g(X)g(Y ) has a legal turn at the concatenation point, as desired. (In general, showing that there are no periodic Nielsen k paths requires showing that there exists some k such that g# (γ) is legal for all immersed edge paths γ, which could take quite a bit more work than in the present example). Next we compute the local Whitehead graphs at p and q. The turns taken by the images of all edges under g are {A, F }, {D, F }, {F , A}, {F , D}, {F , E} Iterating these turns under powers of Dg we obtain: p
:
qY
:
{F , E}
{A, F }
{D, F }
/ {F , D} V
{E, D} V
{F , A} V
These are all the turns taken by iterates of edges under g, so the local Whitehead graphs are p
:
A
F
D
q
:
A
F
D
E
E These two graphs are connected, the transition matrix of g is obviously irreducible, and g has no periodic Nielsen paths, and so the outer automorphism φ represented by g is fully irreducible. Also, g having no Nielsen paths, it follows that φ is not geometric, and the ideal Whitehead graph of φ is just the union of the local stable Whitehead graphs of g. The stable directions of g at q being A, F , D, E, the ideal Whitehead graph of φ is therefore A
D
F
46
E
4
Cutting and pasting local stable Whitehead graphs
Let φ ∈ Out(Fn ) be nongeometric and fully irreducible. Corollary 3.2 shows that choosing a train track representative g : G → G of φ induces a decomposition of each component of the ideal Whitehead graph W (φ), expressing that component as the union of local stable Whitehead graphs pasted together along cut vertices. We refer to this as the local decomposition of W (φ) induced by g. In this section we study how the local decomposition changes under certain moves that alter the train track representative. Our main result, Corollary 4.4, explains how to construct a train track representative so that the local decomposition is as fine as possible, having no cut vertices in any local stable Whitehead graph. To put it another way, every finite graph has a unique decomposition into a union of maximal subgraphs without cut vertices, called the cut point decomposition of the graph, and Corollary 4.4 produces a train track representative that simultaneously realizes the cut point decomposition for all components of W (φ). The local decomposition of W (φ) is intimately connected with Nielsen classes of principal vertices of train track representatives. Given a train track representative g : G → G of φ, while the number of Nielsen classes of principal vertices of g depends only on φ, the number of principal vertices depends on the choice of g : G → G. As Lemma 3.1 and Corollary 3.2 show, for any train track representative g : G → G of φ, the decomposition of W (φ) into local stable Whitehead graphs of g pasted together along cut vertices corresponds to the decomposition of Nielsen classes of principal vertices of g into individual vertices joined by Nielsen paths. First, in Lemma 4.1, we describe a condition which allows us to reduce the number of principal vertices, and a method for doing so — this corresponds to coarsening the local decomposition of W (φ). Then, in Lemma 4.3, we give necessary and sufficient conditions for recognizing when a given irreducible train track map is the product of this method. In this case the method can be reversed to increase the number of principal vertices, refining the local decomposition of W (φ). As an application, in Corollary 4.4 we show that the finest decomposition of W (φ) is realized by maximizing the number of principal vertices of a train track representative; the maximum exists because the number of principal vertices is bounded above by the number of singular leaves of Λ− .
4.1
Pasting local stable Whitehead graphs
Close study of Lemma 3.1 and Corollary 3.2 indicates that we should be able to paste together local stable Whitehead graphs and coarsen the local decomposition of the ideal Whitehead graph, by decreasing the number of principal vertices, or equivalently by eliminating Nielsen paths. The following describes a process for carrying out this elimination; it is a variation on Lemma 3.8 of [BH92]. Lemma 4.1. Suppose that g : G → G is an irreducible train track map, that E1 and E2 are ¯2 is an indivisible periodic Nielsen edges in G with a common terminal vertex and that E1 E path that either has period one or is mapped by f# to its inverse (and so has period two). Let G′ be the graph obtained from G by isometrically identifying E1 with E2 and let p : G → G′ be the quotient map. Then there is an induced irreducible train track map g ′ : G′ → G′ with the following properties:
47
(1) g ′ p = pg. e ′ → T+ are as in Corollary 2.14 then p lifts to p˜ : G e→G e′ e → T+ and fg′ : G (2) If fg : G such that fg = fg′ p˜. (3) If g is a train track representative of φ ∈ Out(Fr ) via a marking ρ : Rr → G, then g ′ is also a train track representative of φ via the marking p ◦ ρ : Rr → G′ . (4) The function p maps the principal vertices of g onto the principal vertices of g ′ , preserving Nielsen classes. This map is one-to-one except that the endpoints of E1 E¯2 are identified to a single point. In this lemma, the local decomposition of W (φ) induced by g is obtained from that induced by g ′ by pasting the local stable Whitehead graphs of g ′ at the initial endpoints of E1 , E2 , identifying the direction of E1 with the direction of E2 . We shall refer to the construction of this lemma as a Nielsen path collapse performed on ¯2 . E1 E Proof. The initial endpoints of E1 and E2 are distinct because φ does not fix the conjugacy class of any basis element. There is a path τ such that g(Ei ) = Eσ(i) τ where i = 1, 2 and σ : {1, 2} → {1, 2}. Denote p(E1 ) = p(E2 ) by E ′ and define g ′ (E ′ ) = E ′ p(τ ). Every other edge of G′ equals p(E) for some edge E of G and we define g ′ (p(E)) = pg(E) which is an immersed path because p restricts to an immersion on legal paths in G. For the same reason (g ′ )k (p(E)) = pg k (E) for all k > 0 which implies that g ′ is an irreducible train track map. e → G e ′ be the lift that fixes all edges that do Item (1) is true by construction. Let p˜ : G ′ not project to E1 or E2 . Then fg and fg p˜ are equivariant, agree on vertices and restrict to isometries on all edges. It follows that fg = fg′ p˜, proving (2). Item (3) is obvious. To prove (4), let wi be the initial point of Ei and let w′ = pw1 = pw2 . The map p is injective on vertices except that pw1 = pw2 . A vertex v of G is g-periodic if and only if pv is g ′ -periodic; for each of the two cases v ∈ {w1 , w2 } and v 6∈ {w1 , w2 } this is obvious, but for slightly different reasons. We now assume that v ∈ G is a principal vertex and prove, by consideration of cases, that pv ∈ G′ is principal. Case 1: if v has at least three Dg-periodic directions then their images under Dv p give at least three Dg ′ -periodic directions at pv. ¯2 or its inverse, then Case 2: if v is the endpoint of a periodic Nielsen path γ other than E1 E pγ = p# γ is a periodic Nielsen path and has pv as an endpoint. Case 3: if v = w1 or w2 then Dg has at least two periodic directions at each of w1 and w2 , and so pv = w′ has at least three periodic directions, namely Dw1 p(E1 ) = Dw2 p(E2 ), one more in image(Dw1 p), and one more in image(Dw2 p). It follows that if v is principal then so is pv. Conversely, every vertex of G′ with at least three periodic directions occurs as in Case 1 or Case 3, and every periodic Nielsen path of g ′ occurs as in Case 2. This proves that v is principal for g if and only if pv is principal for g ′ , and moreover that Nielsen classes are preserved. The statement about local decompositions is obvious by construction. ♦ A Nielsen path collapse is similar to the more familiar operation from [BH92] of doing a fold on a turn of a train track map to produce a new train track map. We note some of the properties of the latter; the proof is similar to Lemma 4.1 but easier.
48
Lemma 4.2. Consider an irreducible train track map g : G → G and oriented edges E1 , E2 ⊂ G with a common terminal vertex so that g(E1 ) = g(E2 ). Let p : G → G′ be the Stallings fold, the quotient map that identifies E1 to E2 . Then there is an induced irreducible train track map g ′ : G′ → G′ that satisfies items (1)–(3) of Lemma 4.1, and such that (4′ ) The function p restricts to a bijection between the principal vertices of g and g ′ preserving Nielsen classes. The local decompositions of the ideal Whitehead graph induced by g and by g ′ are identical. ♦ We say that g ′ is obtained from g by a fold of E1 with E2 .
4.2
Cutting local stable Whitehead graphs
A Nielsen path collapse is reversible, and one can ask how to recognize when a given irreducible train track map g ′ : G′ → G′ is the product of a Nielsen path collapse applied to some g : G → G as in Lemma 4.1. Since the map p : G → G′ restricts to an immersion on leaves of Λ− it induces a map of local Whitehead graphs Dp : Ww → Wp(w) for each vertex w of G. Similarly g : G → G induces a map Dg : Ww → Wg(w) . If wi ∈ G is the initial vertex of Ei and w′ = p(wi ) ∈ G′ then Ww′ decomposes as the union of Dp(Ww1 ) and Dp(Ww2 ) and these two subgraphs of Ww′ intersect in the single vertex determined by E ′ = p(Ei ). This vertex is therefore a cut vertex of Ww′ , meaning that it separates Ww′ . Note also that if g ′ (v ′ ) = w′ for some vertex v ′ 6= w′ then Dg ′ (Wv′ ) is contained in either Dp(Ww1 ) or Dp(Ww2 ). Finally, w′ is a fixed point of g ′ and the action of Dg ′ on Ww′ preserves the decomposition Ww′ = Dp(Ww1 ) ∪ Dp(Ww2 ), fixing or interchanging the two sets Dp(Ww1 ), ¯2 . Dp(Ww2 ) depending on whether g# preserves or reverses E1 E The following lemma states that these necessary conditions are also sufficient. The lemma is stated with sufficient breadth so that the construction may be used as well to reverse a fold. Lemma 4.3. Suppose that g ′ : G′ → G′ is an irreducible train track map, that w′ is a vertex of G′ and that (1) Ww′ is the union of two non-trivial subgraphs X1 and X2 that intersect in a single vertex x. (2) If g ′ (v ′ ) = w′ for some vertex v ′ 6= w′ then Dg ′ (Wv′ ) is contained in either X1 or X2 . (3) If g ′ (w′ ) = w′ then Dg ′ (x) = x and there exists σ : {1, 2} → {1, 2} such that Dg ′ (Xi ) = Xσ(i) . Then there exists an irreducible train track map g : G → G and a pair of edges E1 and E2 in G with a common terminal vertex and with distinct principal initial vertices w1 and w2 such that the quotient map p : G → G′ that identifies E1 to E2 is a homotopy equivalence satisfying: (4) pg = g ′ p. (5) p(w1 ) = p(w2 ) = w′ . 49
(6) p(E1 ) = p(E2 ) = E ′ is the edge in Ww′ corresponding to x. (7) Dp(Wwi ) = Xi for i = 1, 2. Furthermore, if g ′ (w′ ) = w′ then g and the edges E1 , E2 satisfy the hypotheses of Lemma 4.1 ¯2 , whereas if g ′ (w′ ) 6= w′ then and g ′ is obtained from g by a Nielsen path collapse along E1 E ′ g and E1 , E2 satisfy the hypotheses of Lemma 4.2 and g is obtained from g by a fold of E1 with E2 . Proof. For i = 1, 2 denote the edges of G′ that correspond to vertices of Xi other than x by Pi′ . Let u′ be the terminal vertex of the edge E ′ corresponding to x. Define a new graph G from G′ by replacing w′ with a pair of vertices wi , by replacing E ′ with a pair of edges Ei with initial vertex wi and terminal vertex u′ and by reattaching the edges in Pi′ to wi . The map p : G → G′ that identifies E1 to E2 is a homotopy equivalence. We construct g : G → G as follows. If g ′ (w′ ) = w′ then g(wi ) := wσ(i) ; otherwise g(wi ) := p−1 g ′ (w′ ). If v is a vertex of G such that g ′ (p(v)) 6= w′ , then g(v) := p−1 (g ′ (p(v))). If v is neither w1 nor w2 and if g ′ (p(v)) = w′ then g(v) := wi where Dg ′ (Wv ) ⊂ Xi . If E is an edge with endpoints s and t then define g(E) to be the unique path with endpoints g(s) and g(t) such that pg(E) = g ′ p(E). This unique lifting exists because the initial and terminal vertices have already been lifted and the base vertex of each turn in g ′ (p(E)) lifts uniquely: at the only vertex w′ that does not lift uniquely to G the turns of g ′ (p(E)) lift uniquely because they are contained in either X1 or X2 . This completes the definition of g. Properties (4)–(7) are clear from the construction. We must still show that g : G → G is an irreducible train track map. The obvious induction argument shows that pg k = (g ′ )k p for all k ≥ 0. If E is an edge of G then p(E) is an edge of G′ and so (g ′ )k p(E) is an immersed path for all k ≥ 0. It follows that g k (E) is immersed and hence that g is a train track map. Each (g ′ )k (pE) is a subpath of a leaf of the realization of Λ− in G′ . It follows that g k (E) is a subpath of a leaf of the realization of Λ− in G. Example 2.5(1) of [BFH97] implies that g k (E) covers G for all sufficiently large k and hence that g is irreducible. ¯2 is a It remains to prove the “furthermore” clause. If g ′ (w′ ) = w′ then clearly E1 E ¯2 periodic Nielsen path mapped by g# to itself or its inverse, so the Nielsen collapse of E1 E ′ ′ ′ ′ −1 ′ ′ is defined and its product is clearly g . If g (w ) 6= w then g(w1 ) = g(w2 ) = p (g (w )), and clearly g(E1 ) = g(E2 ) is the unique path in G with initial point p−1 (g ′ (w′ )) that projects to g ′ (E); the fold of E1 with E2 is therefore defined and its product is clearly g ′ . ♦
4.3
The finest local decomposition
Here is our main result, in which we describe a process that starts with an arbitrary train track representative and produces one whose local decomposition is finer than any other. We do this only in the rotationless case; a more general result, while interesting, is not necessary for our present purposes, and the technical details would be a distraction. Proposition 4.4. Let φ ∈ Out(Fr ) be nongeometric, fully irreducible, and rotationless. If g : G → G is an irreducible rotationless train track map representing φ with the maximal number of principal vertices then SWv has no cut vertices for any principal vertex v of g, and the local decomposition of W (φ) induced by g is the cut point decomposition.
50
The proof gives a process that will produce such a g starting from any train track representative of φ. The steps in the process involve some folds, some inverse Nielsen path collapses, and some inverse folds. Proof. We begin with some preliminary folding that does not change the number of principal vertices. Recall that a pair of directions d1 and d2 at the same vertex w are in the same gate at w if they are identified by some iterate of Dg. Define the gate excess at w to be the number of directions minus the number of gates. Define the total gate excess to be the sum of the gate excesses over all principal vertices and over all nonprincipal, and hence non-periodic, vertices with three or more gates. If w has two directions d1 , d2 that are identified by Dg itself then, after possibly subdividing G, we can fold the edges corresponding to d1 and d2 to create a new nonprincipal vertex with two gates, and so the total gate excess is reduced as long as w is not a nonprincipal vertex with two gates. Since g is rotationless, this can be applied repeatedly to the principal vertices until each of them has zero gate excess. If the total gate excess is still positive then there exists a nonprincipal vertex w with at least three gates and positive gate excess. It follows that some positive iterate of g takes w to a principal vertex, and so there exists a minimal i ≥ 1 such that g i (w) has zero gate excess. Replacing w by g i−1 (w) we may assume that there are two directions d1 , d2 at w that are identified by Dg itself, and so a subdivision and fold at w reduces the total gate excess. Continuing by induction, we may assume that g has zero gate excess. In particular, for every vertex w which is either principal or has at least three gates, Dg Ww : Ww → Wg(w) is injective. Furthermore, Ww = SWw if w is principal. Assuming now that g has a principal vertex w such that Ww = SWw has a cut vertex, we describe a sequence of inverse folds followed by an inverse Nielsen collapse which refines the local decomposition of W (φ) induced by g. Let X1 and X2 be non-trivial subgraphs of Ww whose union is Ww and whose intersection is {x}. Define V to be the set of vertices v 6= w for which there exists i ≥ 0 so that g i (v) = w and so that Dg i (Wv ) is not contained in either X1 or X2 . Thus Dg i (Wv ) contains x, at least one vertex in X1 \ {x} and at least one vertex in X2 \ {x}. In particular, v has at least three gates. Let X1′ ∪ X2′ be the decomposition of Wv defined by pulling back the decomposition X1 ∪ X2 of Ww via Dg i . It follows, by injectivity of Dg Wv , that X1′ ∩ X2′ is a cut vertex x′ . If u is a vertex such that g(u) = v and such that Dg(Wu ) is not contained in either X1′ or X2′ then u ∈ V . Case 1: V = ∅. In this case the three items in the hypothesis of Lemma 4.3 are satisfied: (1) is obvious, (2) follows from the definition of V and (3) follows from the assumption that φ is rotationless. Applying the lemma, an inverse Nielsen collapse produces an irreducible train track map with more principal vertices than g, refining the local decomposition of W (φ) induced by g. Case 2: V 6= ∅. Since g is rotationless, there exists v ∈ V that is not g(u) for any u ∈ V . Then Lemma 4.3 applies to the decomposition X1′ ∪ X2′ described above: (1) is obvious, (2) follows from the choice of v and our discussion above, and (3) is vacuous. Applying the lemma, an inverse fold produces a new irreducible train track map with the same number of principal vertices and the same local decomposition as g : G → G but with one fewer element in V . Continuing inductively until V = ∅, we reduce to Case 1. ♦
51
5
Weak train tracks
Let φ ∈ Out(Fr ) be nongeometric and fully irreducible with expanding lamination Λ− = Λ− (φ) and attracting Fr tree T+ = T+ (φ). Consider a marked graph G or its universal cover e = T . We say that G or T is a train track for φ if there exists an irreducible rotationless G train track map g : G → G representing an iterate (a positive power) of φ, and we say that G or T arises from g : G → G. We say that G or T is a weak train track for φ if, up to renormalizing T and/or T+ , there exists a Λ− isometry fT : T → T+ . The definitions of train tracks and weak train tracks are evidently invariant under the equivalence relation defining Xr , and so we can speak about a point of Xr being a train track or a weak train track. Since the definition of a weak train track for φ depends only on the lamination Λ− and the tree T+ , and since these two objects are insensitive to iterating φ, a weak train track for φ is the same thing as a weak train track for any iterate of φ. We will therefore replace φ with an iterate whenever it is convenient. Henceforth we shall always fix a normalization of T+ a priori. Having done so, we would like to normalize each weak train track T so that there exists a Λ− isometry fT : T 7→ T+ . As we will see in Section 5.4, there are examples where fT is not uniquely determined by T , although the failure of uniqueness is mild (if Λ− is nonorientable then fT is unique; whereas if Λ− is orientable then fT varies over a space of maps homeomorphic to a closed interval in the real line). It might seem that the normalization of T depends on the choice of fT . The following says that this does not happen: Lemma 5.1. Fix a normalization of T+ . For each weak train track T ∈ Xr there exists a unique normalization of T in Xbr for which there exists a Λ− isometry from T to T+ .
Proof. Consider any normalization of T for which there exists a Λ− isometry fT : T → T+ . Consider the marked graph G = T /Fr . Each leaf L of Λ− is realized by a surjective local isometry ℓ : R 7→ G, and so there is an oriented edge e of G and points r < s ∈ R such that ℓ(r) = ℓ(s) ∈ int(e) and ℓ is orientation preserving near r and near s. Surgery of ℓ produces a loop by restricting ℓ to [r, s] and identifying r and s. If γ ∈ Fr represents this loop then the length of this loop is equal to the translation length of γ acting on T . Moreover, by construction the axis of γ in T is legal with respect to the map fT , and so is mapped by fT isometrically and γ-equivariantly onto the axis of γ in T+ . The translation lengths of γ acting on T and on T+ are therefore equal, thereby uniquely determining the normalization of T . ♦ By Lemma 2.20, every train track for φ is a weak train track for φ. To be more explicit, e is a weak train track, for every train track representative g : G → G of φ, the train track G e → T+ is a Λ− isometry. The main technical result of this section is and the map fg : G Proposition 5.4 which explains how to obtain every sufficiently short weak train track by folding one particular train track. We also explain in Proposition 5.5 how to recognize when a weak train track is a train track. As an application we prove in Corollary 5.6 that the set of weak train tracks for φ is dense in the set of train tracks for φ, as subsets of Xr . Much of what we do in this section is to generalize some of the structure of train tracks, such as the local stable Whitehead graphs and the local decomposition of the ideal Whitehead graph W (Φ), to the context of a weak train tracks. These notions will depend not only on the weak train track T but also the choice of a Λ− isometry fT : T → T+ . We will sometimes 52
emphasize the dependence on fT by referring, for example, to the local decomposition of W (Φ) induced by T relative to fT . As we proceed we will begin to suppress this dependence when fT is clear from the context. On first pass, the reader may wish to assume uniqueness of fT (a perfectly safe assumption when Λ− is nonorientable).
5.1
Local decomposition of the ideal Whitehead graph.
Let T be a weak train track for φ. Choose a Λ− isometry fT : T → T+ . If b is a branch point of T+ and Φ ∈ PA(φ) corresponds to b as in Lemma 2.16, then the realization in T of each leaf ℓ˜ ∈ L(Φ) contains a unique point w that is mapped by fT to b, called the base point of ℓ˜ in T . Fixing w ∈ T , as ℓ˜ varies over the leaves in L(Φ) with base point w, the point w divides the realization of ℓ˜ into two rays called singular rays of T at w. Each singular ray of T at w is mapped by fT to a singular ray of T+ at b. There are only finitely many Fr -orbits of basepoints w so we may equivariantly subdivide T to make each basepoint a vertex. A vertex of T is principal if it is the basepoint of some singular leaf. Base points of singular leaves, singular rays, and principal vertices are all defined relative to fT . For each point v of T , the local Whitehead graph W (v; T ) is the graph whose vertices are directions at v and with an edge connecting two vertices if the turn that they determine is ˜ − . Local Whitehead graphs are defined absolutely, not relative to fT . taken by a leaf of Λ For each principal vertex w of T relative to fT , the local stable Whitehead graph SW (w; T ) is the graph whose vertices are initial directions of singular rays based at w and with an edge ˜− connecting two vertices if the turn that they determine is taken by a singular leaf of Λ realized in T with base point w. If fT (w) = b then fT induces an injection of SW (w; T ) into W (b; T+ ); we use this and Lemma 3.4 to identify SW (w; T ) with a subgraph of W (b; T+ ) and with a subgraph of the component W (Φ) of the ideal Whitehead graph W (φ) where Φ ∈ PA(φ) corresponds to b. The local stable Whitehead graph SW (w; T ) is defined relative to fT , so a more detailed notation which we prefer to avoid would be SW (w; fT ). The following lemma generalizes Lemma 3.2. Lemma 5.2. Suppose that T is a weak train track, fT : T → T+ is a Λ− isometry, b is a branch point of T+ , Φ ∈ PA(φ) is the representative of φ corresponding to b, and {wi } are the principal vertices of T that are mapped by fT to b. Then (1) W (Φ) = ∪SW (wi ; T ). (2) For i 6= j, SW (wi ; T ) ∩ SW (wj ; T ) intersect in at most one vertex. If they do intersect at a vertex P , then P is a cut point of W (Φ), in fact P separates SW (wi ; T ) from S(wj ; T ) in W (Φ). Proof. (1) is immediate from the definitions and from our identification of SW (wi ; T ) with a subgraph of W (Φ). The proof of (2) is the same as the proof of the third item of Corollary 3.2. ♦ As in the case of Corollary 3.2, Lemma 5.2 shows that the choice of a weak train track T and a Λ− isometry fT : T → T+ induces a decomposition of the ideal Whitehead graph W (φ), expressing each component of W (φ) as a union of local stable Whitehead graphs of T pasted together along cut vertices. We refer to this as the local decomposition of W (φ) relative to fT . 53
Given two weak train tracks S, T and Λ− isometries fS : S → T+ , fT : T → T+ , we say that fT splits at least as much as fS if the local decomposition W (φ) = ∪SW (vj ; T ) is at least as fine as the the local decomposition W (φ) = ∪SW (wi ; S), meaning that for each principal vertex vj of T there is a principal vertex wi of S such that SW (vj ; T ) ⊂ SW (wi ; S) where the inclusion takes place in W (φ). To put it another way, a Λ− isometry fT : T → T+ imposes an equivalence relation on fT
the set of singular leaves realized in T , where l1 ∼ l2 if l1 , l2 have the same base points fT
relative to fT . Obviously l1 ∼ l2 if and only if their edges belong to the same component of the local decomposition of W (φ) relative to fT . Thus, to say that fT splits at least as much as fS means that, for all singular leaves l1 , l2 , if the realizations of l1 , l2 in T have the same basepoint relative to fT then their realizations in S have the same basepoint relative to fS . Lemma 5.3. There exists a rotationless irreducible train track map g : G → G representing an iterate of φ such that for every weak train track T and Λ− isometry fT : T → T+ , the Λ− e → T+ splits at least as much as fT . If φ is rotationless then we may choose isometry fg : G g : G → G to represent φ. Proof. After replacing φ with an iterate we may assume that φ is rotationless. Apply Lemma 4.4 to choose a rotationless train track map g : G → G whose local stable Whitehead links have no cut vertices. For each Φ ∈ PA(φ), there are decompositions W (Φ) = e and W (Φ) = ∪SW (wj ; T ) where f e (˜ e has no ∪SW (˜ vi ; G) vi ; G) G vi ) = fT (wj ) = b. Since SW (˜ cut vertices, it must be contained in some single SW (wj ; T ). ♦
5.2
Folding up to a weak train track
We now come to the main technical result of this section. This result says that from any fixed train track we can fold to any sufficiently short weak train track which splits no more than the fixed train track. This is very similar to Lemma 3.16 of [LL04] which combines results from [Lus99]. Proposition 5.4. Let φ ∈ Out(Fn ) be nongeometric and fully irreducible. For any train e → T+ , there exists track representative g : G → G for φ with associated Λ− isometry fg : G ǫ > 0 so that if T is any weak train track for φ, if fT : T → T+ is any Λ− isometry, if e → T+ splits at least as much as fT , and if Len(T ) ≤ ǫ, then there exists a unique fg : G e → T such that fg = fT h. This map h is moreover a Λ− equivariant edge isometry h : G isometry. In this statement, recall our convention of fixing a normalization of T+ , relative to which e and T are normalized, and so Len(T ) is well defined. G
Proof. Uniqueness of h, and the fact that h must be a Λ− isometry, are proved as follows. e For any leaf L of Λ− realized in G, since fg L is an isometry, h L must be an injection, and since h is an edge isometry the map h L must be an isometry onto its image. Moreover, the image h(L) must be the realization of L in T , which we denote LT , proving that h is a Λ− isometry. Since fT LT is an isometry, it follows that h L = (fT LT )−1 ◦ (fg L). This being true for any L, the map h is uniquely characterized. For the rest of the proof we turn to existence of h. 54
There is no loss in assuming that φ is rotationless. After replacing g : G → G by an iterate we may assume that for any vertex v and any direction d at v ∈ G, the vertex g(v) is fixed and the direction Dv g(d) is fixed. We now assume the notation of Figure 2, and we e 0 for G e and fg0 for fg . write G e0 → T will be defined equivariantly, first on the vertices of G e0 , and then The map h : G by linear extension over the edges. With appropriate choice of ǫ we will then prove that fg0 = fT h, from which it will immediately follow that h is edge isometric. e 0 projects to a fixed point of g then there is a natural choice of h(˜ If x ˜∈G x) ∈ T . To e e take care of a general vertex x ˜ ∈ G0 , we first push x ˜ forward via g˜1,0 from G0 to a point e 1 which does project to a fixed point, and which now has a natural choice of image y˜ ∈ G e0 , we shall define h(˜ in T . In other words, for each vertex x˜ of G x) = h1 g˜1,0 (˜ x) where h1 is e1 as follows. defined on y˜ = g˜1,0 (˜ x) ∈ G e1 If y˜ is principal then it is the basepoint of the realization of some singular leaf ℓ˜ in G ˜ and we define h1 (˜ y ) to be the basepoint of the realization of ℓ in T . This is independent of e 0 , and so also G e 1 , splits the choice of ˜l, and therefore well defined, by the hypothesis that G at least as much as T . If y˜ is not principal then there are exactly two fixed directions at ˜ − by Lemma 2.17. There is a lift of g that fixes y˜ and y˜ and they determine a leaf ℓ˜ of Λ ˜ = ℓ. ˜ The homeomorphism if Φ is the automorphism corresponding to that lift then Φ# (ℓ) ˜ expanding the metric by λ(φ), and so has a Φ+ : T+ → T+ of Corollary 2.21 preserves ℓ, unique fixed point c in the realization of ℓ˜ in T+ . Define h1 (˜ y ) to be the unique point in the realization of ℓ˜ in T that is mapped to c by fT . e0 , we now extend it linearly to all of G e0 . Choose Having defined h on the vertices of G ǫ > 0 to be less than half the length of the shortest edge in G1 . We show next that for any e of G e 0 , the path h(E) e is legal with respect to fT . edge E e For i = 0, 1, let y˜i = g˜1,0 (˜ Let x ˜0 and x ˜1 be the endpoints of E. xi ). There is a lift of g that e fixes y˜i and that fixes the direction at y˜i determined by g˜1,0 (E); let Φi be the automorphism corresponding to this lift of g. We may choose the leaf ℓ˜i used in the definition of h1 (˜ yi ) e To be precise, ℓ˜i is the leaf through to ‘contain’ the direction at y˜i determined by g˜1,0 (E). e1 is fixed by (Φi )# . The leaf ℓ˜i contains a non-trivial segment of y˜i whose realization in G e incident to y˜i , and its realization in T contains h1 (˜ g˜1,0 (E) yi ) = h(˜ xi ). Let ℓ˜2 be any leaf ˜ e e e e whose realization in G0 contains E, and so the realization of ℓ2 in G1 contains g˜1,0 (E). e Let z˜0 , z˜1 be the midpoints of the initial and terminal edges of g˜1,0 (E) respectively. For i = 0, 1, the realizations of ℓ˜i and ℓ˜2 in T+ contain an interval in T+ of radius greater zi ). Since T has length at most ǫ, the number ǫ is a cancellation than ǫ centered on fg1 (˜ constant (Lemma 2.5) for fT : T → T+ . It follows that the intersection of the realizations of zi ). One can ℓ˜i and ℓ˜2 in T contains an open arc about a point w ˜i that maps by fT to fg1 (˜ therefore travel along a path in T that starts from h(˜ x0 ) = h1 (˜ y0 ), goes along the realization of ℓ˜0 to w ˜0 , then along the realization of ℓ˜2 to w ˜1 , and then along the realization of ℓ˜1 to h1 (˜ y1 ) = h(˜ x1 ), without ever taking an illegal turn. This legal path is therefore the unique e embedded arc in T from h(˜ x0 ) to h(˜ x1 ), which by definition is h(E). e e 0 → T+ is an x) for all vertices x ˜ of G0 . Thus fT h : G By construction, fg0 (x) = fT h(˜ equivariant map that agrees with fg on vertices and is linear on edges. It follows that fg0 = fT h as desired. Also, since fg0 and fT are both isometric on edges, it follows that h
55
♦
is isometric on edges.
5.3
Comparing train tracks to weak train tracks
There are two moves one can perform on a train track map g : G → G representing an iterate of φ to produce a weak train track that is not a train track. The first move is to partially fold a pair of edges to create a vertex that is not preperiodic for the induced map. In this case, although there is an induced homotopy equivalence on the quotient space, this map does not take vertices to vertices no matter how the quotient space is subdivided as a CW complex. The second move is to collapse a pre-Nielsen path by a folding operation that is analogous to the one used in Lemma 4.1 without also collapsing the Nielsen path to which it maps. In this case there is no induced map on the resulting quotient space. The next proposition shows that these are the only possibilities, and the following corollary uses this to prove that train tracks are dense in the set of weak train tracks. We say that c ∈ T+ is a periodic point if there is an automorphism Φ representing a rotationless iterate of φ such that c ∈ Fix(Φ+ ). Every branch point of T+ is a periodic point by Lemma 2.16. Also, if g : G → G is a train track representative of some rotationless iterate e projects to a periodic point of g then fg (x) of φ then Theorem 2.15 guarantees that if x ∈ G is a periodic point of T+ , proving the necessity of item (1) in the following: e be a weak train track and fT : T → T+ a Λ− isometry. The folProposition 5.5. Let T = G lowing are necessary and sufficient conditions for the existence of a train track representative g : G → G of φi for some i, such that fT = fg : (1) For every vertex w of T , fT (w) ∈ T+ is a periodic point. (2) For every vertex y of T , if fT (y) = b is a branch point of T+ then there exists a principal vertex w of T such that DfT (W (y; T )) ⊂ DfT (SW (w; T )) ⊂ W (b; T+ ). Proof. For the necessity of item (2), choose i ≥ 1 so that g i maps every vertex to a fixed point of g i and Dg i maps every direction at a vertex to a fixed direction. Referring to the notation e=G e0 such that y and w have of Figure 2, it follows that there is a principal vertex w of T = G ei under g˜i,0 = g˜i,i−1 ◦ · · · ◦ g˜1,0 , and Dgi,0 (W (y; T )) ⊂ SW (wi ; G e i ). the same image wi ∈ G Since Dg i is the identity on the set of fixed directions at fixed vertices of g i , the map D˜ gi,0 e induces an isomorphism between SW (w; T )) and SW (wi ; Gi ). Applying the “chain rule” DfT = Dfg0 = Dfgi ◦ Dfgi,0 , it follows that DfT (W (y; T )) ⊂ DfT (SW (w; T )) ⊂ W (b; T+ ). For sufficiency, let R be the set of rotationless train track maps g : G → G representing e → T such that h is simplicial an iterate of φ for which there exists an equivariant map h : G (vertices map to vertices and edges map to edges) and fg = fT h. We first show that R = 6 ∅. By Lemma 5.3 there exists a rotationless train track map e → T+ splits g : G → G representing an iterate of φ such that the associated train track fg : G as least as much as fT : T → T+ . As there are only finitely many indivisible pre-Nielsen paths for g, we may assume, after replacing g with an iterate, that if µ is a pre-Nielsen path for g, then g# (µ) is a Nielsen path for g. Choose ǫ as in Proposition 5.4 applied to e → T+ . Replacing G with φ−n G has the effect of replacing the constant ǫ with λn ǫ. fg : G Choosing n sufficiently large, we may therefore assume that Len(T ) ≤ ǫ and hence that there e → T such that f e = fT h. Up to the action of Fr , there are only is an equivariant map h : G G 56
e that map via h to vertices of T . Suppose that w is a vertex of T , finitely many points of G e and that h(˜ that x ˜ is a point of G x) = w. Let x be the image of x˜ in G. After replacing φ g of g(x) such that with an iterate, (1) and Theorem 2.15(1) imply the existence of a lift g(x) g fg (g(x)) = fg (˜ x). Corollary 2.14 implies that for some m ≥ 0 either g m (x) is the endpoint of Nielsen path or g m+1 (x) = g m (x). In either case g m (x) ∈ Fix(g). We may therefore subdivide along the g-orbit of x to make x and hence x ˜ a vertex. After finitely many such subdivisions, the map h is simplicial. We have now shown that R = 6 ∅. e → T descends to a simplicial homotopy For each g : G → G in R the associated map h : G ¯ : G → H where the marked graph H is the quotient space of T by the action equivalence h ¯ preserves marking. If h ¯ is a homeomorphism then T = G e and fT = fg , and so of Fr , and h we are done. Otherwise, there exist edges E1 and E2 of G, oriented so as to have a common ¯ Arguing by cases that terminal endpoint, such that E1 , E2 are identified by the map h. depend on the nature of E1 , E2 , we will show that there is an element g ′ : G′ → G′ of R such that G′ has fewer edges than G. The proof then concludes by induction. ¯2 is an indivisible Nielsen path and Lemma 4.1 There are two easy cases. In case 1, E1 E ′ ′ ′ k k produces the desired g : G → G . In case 2, there exists k ≥ 1 such that g# (E1 ) = g# (E2 ), k ′ ′ ′ and replacing g by g , Lemma 4.2 produces the desired g : G → G . If neither case 1 nor case 2 applies to the pair E1 , E2 , then Corollary 2.14 implies, after ¯2 ) is an indivisible Nielsen path. replacing g with an iterate if necessary, that σ := g# (E1 E e ei with a common Lift σ to a path σ ˜ ⊂ G with endpoints v˜1 and v˜2 . There are lifts E terminal endpoints and with initial endpoints x ˜i such that (in the notation of Figure 2) g˘0 (˜ xi ) = v˜i . Let y = h(˜ x1 ) = h(˜ x2 ) and wi = h(˜ vi ). In particular, the endpoints v1 and v2 of σ are principal. Lemma 2.11 and Lemma 2.16 imply that fT (y) = fg (˜ xi ) = fg (˜ vi ) = e ⊂ DfT (W (y; T )) fT (wi ) = b ∈ T+ is a branch point. By construction, Dfg (W (˜ xi ; G)) e ⊂ Dfg (SW (˜ e ⊂ DfT (SW (wi ; T )). For each i = 1, 2, the graph and Dfg (W (˜ xi ; G)) vi ; G)) DfT (W (y; T )) therefore intersects DfT (SW (wi ; T )) in at least one edge. Applying (2) and Lemma 5.2 it follows that DfT (W (y; T )) ⊂ DfT (SW (wi ; T )), and so DfT (SW (w1 ; T )) and DfT (SW (w2 ; T )) intersect each other in at least one edge. Another application of Lemma 5.2 implies that w1 = w2 . In particular, letting (E1′ , E2′ ) be the illegal turn in σ, we have that ¯h(E ′ ) = h(E ¯ ′ ). Replacing E1 and E2 with E ′ and E ′ we are in case 1 if σ = E ′ E ¯′ 1 2 1 2 1 2 and in case 2 otherwise. ♦ Corollary 5.6. For any nongeometric, fully irreducible φ, a dense set of weak train tracks are train tracks. Proof. A point in T is said to be preperiodic if its image under fT is a periodic point in T+ . We define a complexity C(T ) for weak train train tracks T as follows. First define the ‘valence complexity’ V C(T ) to be the non-increasing sequence of integers with one entry ki for each Fr -orbit of vertices wi of valence ki ≥ 3. Then define the ‘aperiodic complexity’ AC(T ) to be the number of orbits of vertices in T that are not preperiodic. Finally, define C(T ) := (V C(T ), AC(T )) ordered lexicographically. Thus C(T ′ ) < C(T ) if V C(T ′ ) < V C(T ) or if V C(T ′ ) = V C(T ) and AC(T ′ ) < AC(T ). It suffices to show that if T is not a train track then there is an arbitrarily close weak train track T ′ with C(T ′ ) < C(T ).
57
Let g : G → G be a rotationless irreducible train track map representing an iterate of φ. e there is a dense set of points that project to periodic points in G, and so for In the tree G ˜ The set of preperiodic e − realized in T+ the set of periodic points is dense in ℓ. any leaf ℓ˜ of Λ points in T is therefore dense. Suppose that w is a vertex of T that is not preperiodic. Then fT (w) is not a branch point, and so the local Whitehead graph W (fT (w); T+ ) is a single edge. Consider the map DfT : W (w; T ) → W (fT (w); T+ ). If w has valence three then there are directions d1 6= d2 at w which have the same image under DfT . Folding these directions up to nearby points whose fT -image is periodic produces a new valence three vertex whose fT -image is preperiodic and reduces w to valence two; we then remove w from the list of vertices of T . This does not change V C(T ) and lowers AC(T ). If the valence of w is higher than three then perform a similar small fold to reduce V C(T ). We may therefore assume that each vertex of T is preperiodic. By this same folding argument we may assume that if the image DfT (W (w; T )) is not a single edge of W (fT (w); T+ ) then the restricted map DfT : W (w; T ) → W (fT (w); T+ ) is an injection. Proposition 5.5 implies that if T is not a train track, then there is a vertex y of T such that fT (y) = b is a branch point of T+ but DfT (W (y; T )) ⊂ W (b; T+ ) is not contained in DfT (SW (w; T )) for any principal vertex w of T . In particular, DfT (W (y; T )) contains at least three vertices of W (b; T+ ), so DfT (W (y; T )) is not a single edge of W (fT (w); T+ ), and so DfT is injective on W (y; T ). Lemma 5.2 implies that W (y; T ) has a cut vertex d, and so W (y; T ) is the union of two subcomplexes X1 and X2 , each containing d properly, such that X1 ∩ X2 = {d}. We now perform a small splitting very similar to the one in the proof of Lemma 4.3. Let E be the edge in the link of y that corresponds to d and let Pi be the edges in the link of y that correspond to the vertices of Xi \{d}. After subdividing T at preperiodic points, we may assume that the length of E is arbitrarily small. Define a new tree T ′ from T by equivariantly replacing y with a pair of vertices yi , by equivariantly replacing E with a pair of edges Ei with initial vertex yi and terminal vertex equal to the terminal vertex of E and by equivariantly reattaching the edges in Pi′ to yi . The map fT ′ defined in the obvious way makes T ′ a weak train track of lesser complexity. ♦
5.4
Rigidity and irrigidity of Λ− isometries
In this section we study uniqueness of Λ− isometries. Given a marked graph G ∈ Xbr with e let I(T, T+ ), denote the space of Λ− isometries T 7→ T+ equipped with universal cover T = G the topology of pointwise convergence. Assuming that T is a weak train track, in other words I(T, T+ ) is nonempty, we wish to know whether this space has more than one point, and if it does then we wish to have some description of its structure. The answer, as it turns out, depends on orientability properties of Λ− : if Λ− is nonorientable then I(T, T+ ) is a single point; and if Λ− is orientable then I(T, T+ ) is a compact interval, possibly degenerating to a point. Recall that D2 ∂Fr is the set of ordered pairs of elements of ∂Fr , that GFr is the space of unordered pairs of elements of ∂Fr and that Λ− is a closed Fr -equivariant subset of GFr . An orientation of Λ− is a lift of the inclusion map Λ− ֒→ GFr to a continuous, Fr -equivariant map o : Λ− → D2 ∂Fr . Equivalently, an orientation of Λ− is a choice of orientation on each leaf of Λ− with the following consistency condition: for some, and hence every, Fr tree T ,
58
there is a constant C so that if LT and L′T are the realizations in T of oriented leaves L and L′ of Λ− and if LT ∩ L′T has length at least C then the orientations of LT and L′T agree on LT ∩ L′T . If Λ− has an orientation then it is orientable. Since Λ− is minimal, there are covering translations γ ∈ Fr such that L′T ∩ γ · LT is arbitrarily long. It follows that an orientation on Λ− is determined by an orientation of any of its leaves and hence that Λ− has either zero or two orientations. An orientation of Λ− in T is a choice of orientation on each leaf of Λ− with the feature that the orientations on LT and L′T agree on LT ∩L′T whenever LT ∩L′T contains a nontrivial interval. An orientation of Λ− in T therefore induces a well defined Fr -equivariant orientation on each edge of T . Recall from Lemma 2.20 that each leaf of Λ− has a well defined realization in T+ . Thus this definition works equally well to define an orientation of Λ− in T+ . It is obvious that if Λ− is orientable in some T then it is orientable. We have the following partial converse. Lemma 5.7. If Λ− is orientable then it is orientable in T+ and is orientable in any weak train track T . e be the universal cover of Proof. Let g : G → G be a train track representative of φ, let G e→G e of g. Since gˆ : ∂Fr → ∂Fr G and let o be an orientation on Λ− . Choose a lift g˘ : G is continuous and Fr -equivariant, it acts on the set of orientations of Λ− and this action is independent of the choice of lift g˘. After replacing g with g 2 if necessary, we may assume that gˆ preserves o. Given L and L′ so that LGe ∩ L′Ge is a nontrivial interval, choose n > 0 so that g˘n (L′Ge )∩ g˘n (LGe ) has length greater than the constant C associated to o in the definition of an orientation. The orientations of g˘n (L′Ge ) and g˘n (LGe ) agree on g˘n (L′Ge ) ∩ g˘n (LGe ) which implies that the orientations of L′Ge and LGe agree on L′Ge ∩ LGe . This proves that Λ− is e orientable in G. If LT+ ∩ L′T+ is a non-trivial interval then there is a train track representative g : G → G e so that L and L′ intersect in a non-trivial interval. By our previous with universal cover G e G
e G
e → T+ preserves argument, the orientations of LGe and of L′Ge agree on LGe ∩ L′Ge . Since fg : G orientation on the leaves of Λ− , it follows that the orientations of LT+ and L′T+ agree on LT+ ∩ L′T+ . This proves that Λ− is orientable in T+ . If T is a weak train track then leaves of Λ− that intersect in T also intersect in T+ . An orientation of Λ− in T+ therefore pulls back to an orientation on Λ− in T . ♦
The next theorem gives an abstract description of the set I(T, T+ ). An orientation o of Λ− induces orientations of Λ− in T and in T+ and each element f ∈ I(T, T+ ) preserves the induced orientations. It follows that for each vertex v of S, if e1 , e2 are two edges at v in the same gate with respect to f then e1 , e2 point in the same direction at v: either both point towards v in which case that gate is said to be a negative gate at v; or both point away from v in which case the gate is said to be positive. e and fix f ∈ Theorem 5.8. Fix a weak train track G ∈ Xbr with universal cover T = G, I(T, T+ ). (1) If Λ− is not orientable in G then f is the unique element of I(T, T+ ). (2) If Λ− is orientable in G then: 59
(a) There exists a compact subinterval ∆ of the real line, and a homeomorphism ∆ 7→ I(T, T+ ). We denote this homeomorphism t 7→ ft . (b) For any leaf L of Λ− and for any x ∈ LT+ , the map ∆ 7→ T+ given by t 7→ ft (x) is an orientation preserving isometry onto a segment of the leaf LT+ . (c) Given f ∈ I(T, T+ ), the following hold: i. f is an interior point of I(T, T+ ) if and only if every vertex of T has exactly one positive and one negative gate with respect to f . ii. f is the positive (resp. negative) endpoint of I(T, T+ ) if and only if some vertex of T has more than one positive (resp. negative) gate with respect to f . iii. f is the unique point of I(T, T+ ) if and only if there exist vertices v, w ∈ T such that v has at least two positive gates with respect to f and w has at least two negative gates with respect to f . Even when Λ− is orientable, this lemma leaves open the possibility that f is unique, that is, the interval ∆ degenerates to a point. See below for examples where ∆ is nondegenerate and where it is degenerate. In item (2a) of this theorem, we use the topology of pointwise convergence on I(T, T+ ), but as should be clear from the proof this is equivalent to the Gromov topology on maps (for the definition of this topology see Section 6.4, proof of step 3(d)). Proof. Consider two Λ− isometric maps f, f ′ : T → T+ . For each oriented leaf L of Λ− , let LT ⊂ T and LT+ ⊂ T+ denote the realizations of L as oriented, bi-infinite geodesics. Each of the restricted maps f, f ′ : LT → LT+ is an orientation preserving isometry, and so these two maps differ by a real number δ(f, f ′ ; L), meaning that the map f ′ : LT → LT+ is obtained by postcomposing the map f : LT → LT+ by the translation of LT+ that displaces each point by the amount δ(f, f ′ ; L) (remember that LT+ is oriented). Next we study how δ(f, f ′ ; L) depends on L. Consider another oriented leaf L′ of Λ− . If LT ∩ L′T = J is a nondegenerate segment then the restriction to J of the map f : LT → LT+ equals the restriction to J of the map f : L′T → L′T+ , and the same is true for f ′ . It follows that if LT and L′T induce the same orientation on J then δ(f, f ′ ; L) = δ(f, f ′ ; L′ ), whereas if LT and L′T induce opposite orientations on J then δ(f, f ′ ; L) = −δ(f, f ′ ; L′ ). Note also that δ(f, f ′ ; L) is invariant under the action of Fr on oriented leaves. Suppose now that Λ− is nonorientable in T . Then for any oriented leaf L of Λ− there exists a covering transformation g ∈ Fr such that LT ∩ g · LT is a nondegenerate segment J and the orientations induced on J by LT and g · LT are opposite. It follows that δ(f, f ′ ; L) = δ(f, f ′ ; g · L) = −δ(f, f ′ ; L) and so δ(f, f ′ ; L) = 0 implying that f LT = f ′ LT . Since this is true for every leaf L, and since realizations of leaves of Λ− cover T , it follows that f = f ′ . Suppose now that Λ− is orientable in T , and fix an orientation. If L, L′ are any two oriented leaves then there exists g ∈ Fr such that g · LT ∩ L′T = J is a nondegenerate segment and the orientations on J induced by g · LT and L′T agree. it follows that δ(f, f ′ ; L′ ) = δ(f, f ′ ; g · L) = δ(f, f ′ ; L) 60
This shows that δ(f, f ′ ; L) is a constant independent of L, which we denote δ(f, f ′ ). The map f ′ is therefore determined by f and the constant δ(f, f ′ ): for each x ∈ T , pick any leaf L whose realization in T contains x, and then f ′ (x) is obtained by displacing f (x) along LT+ by the amount δ(f, f ′ ). Fix f ∈ I(T, T+ ). Let ∆ be the set of all real numbers of the form δ(f, f ′ ) for some ′ f ∈ I(T, T+ ). Each d ∈ ∆ equals δ(f, f ′ ) for a unique f ′ ∈ I(T, T+ ) and we define fd = f ′ . To complete the proof of (2a) we must show that ∆ is a closed, bounded interval; continuity follows by noticing that for any d, d′ ∈ ∆ and any x ∈ T , the distance in T between fd (x) and fd′ (x) is exactly |d − d′ |. Item (2b) follows from item (2a) and the definition of δ(f, f ′ ). To show that ∆ is connected, given a Λ− isometry f ′ : T → T+ , and given a number d between 0 and δ(f, f ′ ) it suffices to exhibit a Λ− -isometry f ′′ : T → T+ such that δ(f, f ′′ ) = d. For x ∈ T the segment f (x)f ′ (x) in T+ has length δ(f, f ′ ), and this segment is oriented with initial point f (x) when δ(f, f ′ ) > 0 and terminal point f (x) when δ(f, f ′ ) < 0. Now define f ′′ (x) to be the point on this segment at displacement d from f (x). Boundedness of ∆ follows from the observation that there exist leaves L, L′ ⊂ Λ− such that LT ∩ L′T 6= ∅ and LT+ ∩ L′T+ is bounded. The fact that ∆ is closed follows by application of the Ascoli-Arzela theorem together with the observation that the compact-open limit of a sequence of Λ− isometries T 7→ T+ is a Λ− isometry, and the fact that on I(T, T+ ) the compact open topology agrees with the pointwise convergence topology. Item 2(c)iii follows from items 2(c)i and 2(c)ii, to whose proofs we now turn. Suppose first that every vertex v ∈ T has a unique positive gate with respect to f , so there is a positive length oriented segment Jv in T+ with initial endpoint f (v) such that every edge in T with initial endpoint v has an initial subsegment mapping isometrically to Jv . There being only finitely many orbits of vertices, we can truncate to assume that all the segments Jv have the same length ǫ > 0. We can then alter f to a new Λ− isometry, by moving each vertex v along Jv to the positive endpoint of Jv , and extending by an isometry on each edge of T . Applying item (2b) it follows that f is not the positive endpoint of I(T, T+ ). Conversely, suppose that there exists a vertex v ∈ T having two or more positive gates, and so there exist edges e, e′ of T each with initial endpoint v such that f (e), f (e′ ) have e − whose realizations in T contain e, e′ , and distinct germs at f (v). Choose leaves L, L′ ⊂ Λ hence each passes through v. The realizations of L, L′ in T+ , while passing through f (v), are disjoint in the positive direction beyond f (v). By item (2b) it follows that f is the positive endpoint of I(T, T+ ). Combining this argument with that of the last paragraph, and repeating these arguments with orientation reversed, items 2(c)i and 2(c)ii follow. ♦
5.5
Examples of exceptional weak train tracks
We describe first a weak train track T for which the space of Λ− isometries I(T, T+ ) is not just a single point. Then we describe a weak train track T that is not a train track. Example: a weak train track T with nontrivial I(T, T+ ). Let a, b and c be generators of F3 . Consider φ ∈ Out(F3 ) represented by the following positive automorphism Φ, shown
61
together with its inverse: a Φ: b c
7 bacaaca → 7→ baca 7→ caaca
a Φ−1 : b c
7→ ¯ba¯ba¯ c 7→ b¯ ab 7→ c¯ abc¯ ab¯ ab
As realized on the three petalled rose, these formulas each define train track maps representing φ and φ−1 , respectively. The automorphism Φ being positive, its transition matrix equals the matrix of its abelianization. Direct calculation shows that no eigenvalue of the abelianization of φ is a root of unity so φ is fully irreducible by applying the following, whose proof is given below. Lemma 5.9. If φ ∈ Out(F3 ) is reducible then at least one eigenvalue of the abelianization φ# is a root of unity. Direct calculation of the Perron-Frobenius eigenvalues of the transition matrices of Φ and Φ−1 also show that λ(φ) > 6 and λ(φ−1 ) < 5, so φ is not geometric. Let g : G → G be the affine train track map representative of φ determined by the above formula for Φ. Let λ be the expansion factor, Λ− the expanding lamination, and T+ the attracting tree for φ. When the three petalled rose G is oriented so that its three oriented e It is clear that edges represent a, b and c then it induces an orientation on Λ− in T = G. at the unique vertex of G the map g has exactly two positive gates {a, b} and {c} and one negative gate {¯ a, ¯b, c¯}. Since each of these gates contains a fixed direction, these are also the gates for fg : Te → T+ . Theorem 5.8 item 2(c)ii implies that fg is the positive endpoint of I(T, T+ ) but is not the negative endpoint. Now we shall slide the map g backwards to locate the negative endpoint of I(T, T+ ). Looking at the formula for Φ we see that the g image of each edge ends with the path aca. Postcomposing Φ with the inner automorphism w 7→ (aca)w(aca)−1 we obtain a 7→ acabaca ′ Φ : b 7→ acab c 7→ acaca This formula defines a train track representative g ′ : G → G of φ supported on the same marked graph G as the train track representative g. Notice now that g ′ has one positive gate {a, b, c} and two negative gates {¯ a, c¯} and {¯b} at its unique vertex. As before each gate contains a fixed direction so these are also the gates of fg′ . Thus fg′ is the negative endpoint of I(T, T+ ). To complete the example we give: Proof of Lemma 5.9. Since φ is reducible there is a free factor F of rank one or of rank two whose conjugacy class is invariant under an iterate of φ. After passing to a power of φ, the conjugacy class of F becomes φ-invariant. If F has rank one then the conjugacy class of a generator of F is φ-invariant, and the image of this generator in Z3 is an eigenvector with eigenvalue one. We may therefore assume that F has rank two. Lemma 3.2 of [BFH00] implies that, after passing to a power of φ, there exist Φ ∈ Aut(F3 ) representing φ and generators a, b and c of F3 such that a and b generate F and such that Φ(c) = ucv for some 62
u, v ∈ F . Letting x be the abelianized image of c and Z 2 the abelianized image of F , we have φ# (x) = x + s for some s ∈ Z2 . If 1 is an eigenvalue for the restriction of φ# to Z2 , we are done. Otherwise there exists t ∈ Z2 such that φ# (t) = t − s and x + t is an eigenvector with eigenvalue one. ♦ Example: a weak train track that is not a train track. We describe an example of a fully irreducible φ ∈ Out(Fr ) and a weak train track H for φ that is not a train track for any iterate of φ. We shall also sketch the proof that H is not even the domain of a nonsimplicial train track representative of any iterate of φ. Start with the outer automorphism φ ∈ Out(F3 ) and the train track representative g : G → G that is described in Section 3.4. The local Whitehead graphs at the two vertices p, q of G, and the stable Whitehead graph at q, are also described. Note that in both the local Whitehead graph and the stable Whitehead graph at q, the vertex D is a cut vertex. We construct H by a small splitting of G at q in the direction D, as was done in Lemma 4.3 and Corollary 5.6. The effect of this is to split the vertex q into two vertices q1 , q2 . An initial segment of the edge D is split into two new edges D1 , D2 , with the tail of Di incident to qi , and with the heads of D1 , D2 incident to a new valence 3 vertex r. The remainder of D becomes an edge D3 from r to p. The other four ends A, E, E, F of G at q are detached from q and reattached as follows: A, E, and F are reattached to q1 , and E is reattached to q2 . A F q1 o @@ @@ D1 @@ E @@ /r q2
@p D3
D2
To be precise about metric and marking, the edges lengths of A, E, F in H are the same as in G, the LenH (D1 ) = LenH (D2 ) = ǫ > 0, and LenH (D3 ) = LenG (D) − ǫ. Let h : H → G be the edge isometry that is the “identity” in A, E, F , and that maps Di ∗ D3 isometrically to D for i = 1, 2. The marking of G composed with a homotopy inverse of h defines the marking of H. ˜: H e → T+ , and let h e →G e be an F3 equivariant lift of h. Consider the Λ− isometry fg : G ˜: H e → T+ is a Λ− isometry, To show that H is a weak train track we show that fh = fg ◦ h ˜ to and to do this we must check that each leaf L of Λ− realized in G lifts isometrically via h e This follows immediately from the observation that in the local the realization of L in H. Whitehead graph of G at q (see Section 3.4), the four edges AF , EF , DF , DE lifts to edges AF , EF , D1 F at q1 and D2 E at q2 . e → T+ , by applying Theorem 5.8. We prove that fh is the unique Λ− isometry from H Since Λ− is oriented we must check condition 2(c)iii in the theorem. Note that the directions A and D of G at p are in distinct negative gates, because A and D are mapped by Dg to e of the vertex p E and F at q, and the latter are fixed by Dg; it follows that any lift to H of H has two negative gates, containing lifts of the directions A and D respectively. Note also that in G the directions A and D at q are in distinct positive gates, because they are
63
e of q1 has two positive gates, containing lifts of the fixed by g; it follows that any lift to H directions A and D1 respectively. To prove that H is not a train track of any iterate of φ, by uniqueness of the Λ− isometry e be a fh it suffices to prove that fh does not satisfy item (2) of Proposition 5.5. Let p˜ ∈ H ˜ p) is a lift of p ∈ G. We know that there is a lift qe ∈ G of q ∈ G such lift of p ∈ H, and so h(˜ ˜ p)) = fg (˜ that b = fg (h(˜ q ) is a branch point of T+ and Dfg maps the four stable directions e ¯ ¯ to the four directions of T+ at b, which shall have the same of G at q˜, named {A, F , D, E} ˜ p) taking the three directions {A, ˜ p) to the ¯ F, D} ¯ at h(˜ names. We also know that fg maps h(˜ ¯ ¯ ¯ F, D} ¯ at three directions {F , D, E} at b. It follows that Dfh maps the three directions {A, ¯ ¯ e p˜ to the three directions {F , D, E} at b. The only principal vertices in H that map to b via ˜ q2 ) = q˜. However, the image under Dfh of fh are the lifts q˜1 , q˜2 of q1 , q2 such that ˜h(˜ q1 ) = h(˜ e at q˜i is the set {A, F¯ , D} at b when i = 1, and the set {D, E} ¯ the set stable directions of H ¯ as is required by item (2) of when i = 2, and neither of these two sets contains {F¯ , D, E}, Proposition 5.5. We also sketch the proof that H is not the domain of any nonsimplicial train track representative of any iterate of φ, meaning a homotopy equivalence g : H → H that represents φ such that for each edge e of H and each i ≥ 1 the map g int(e) is a local homothety with stretch factor λ(φ)i . To prove this requires generalizing several results about train track representatives to the context of nonsimplicial train track representative. Suppose that there is a nonsimplicial train track representative g : H → H of some iterate of φ. We can lift g to define maps as denoted in Figure 2 (replacing the symbol G by H), and then Theorem 2.13 generalizes to produce T+ as a direct limit, and to show that the unique Λ− e → T+ is the direct limit map denoted by fg in Figure 2. Now the argument isometry fh : H given in Proposition 5.5 for the necessity of item (2) generalizes in a straightforward manner. Finally, we have already proved that fh fails to satisfy item (2).
64
6
Topology of the axis bundle
For this section fix a nongeometric, fully irreducible φ ∈ Out(Fr ) with Λ− = Λ− (φ) and T+ = T+ (φ), and fix a normalization of T+ . Having so far proved the equivalence of two of the three proposed definitions of the axis bundle — the set of weak train tracks (2), and the closure of the set of train tracks (3) — in this section we work with definition (2), in terms of which we prove a version of Theorem 1.2. Let A ⊂ Xr be the set of all weak train tracks of φ. By Lemma 5.1, each T ∈ A has a normalization Tb ∈ Xbr well-defined by requiring existence of a Λ− isometry Tb 7→ T+ , and so Len(T ) = Len(Tb/Fr ) ∈ (0, +∞) is well-defined. This map satisfies the equivariance condition Len(T · φi ) = λ(φ)−i Len(T ) for any i ∈ Z because from any Λ− isometry Tb 7→ T+ we obtain a Λ− isometry Tb · φi 7→ T+ · φi = λ(φ)i T+ which, after renormalizing, becomes a Λ− isometry (λ(φ)−i Tb) · φi → T+ . Let Ab ⊂ Xbr be the set of normalized weak train tracks; projection Xbr 7→ Xr restricts to a bijection Ab 7→ A. Abusing notation we often use the same symbol to represent a weak train track and its normalization, although we will the symbol b to maintain the distinction when needed. Here is the main result of this section, a major step in the proof of Theorem 1.2: Theorem 6.1. The map Len : A → (0, +∞) is a proper homotopy equivalence. Outline. Section 6.1 proves basic topological facts about A, including continuity of the normalization map A 7→ Ab and the resulting continuity of the length map Len : A → (0, +∞). b that will Section 6.2 proves a technical result, strengthening the topological properties of A, be useful in the proof of Theorem 6.1. Section 6.3 proves that Len : A → (0, +∞) is a proper map, using folding methods combined with Proposition 5.4. Section 6.4 proves Theorem 6.1 using Skora’s method [Sko] for studying homotopy types of spaces of trees (see also [Whi93], [Cla05], [GL05]).
6.1
Continuity properties of the normalized axis bundle
We prove here several facts about the set Ab and related facts about the length map.
Lemma 6.2.
(1) A is a closed subset of Xr and Ab is a closed subset of Xbr .
(2) The normalization map A → Ab ⊂ Xbr is a homeomorphism.
(3) The map Len : A → (0, +∞) is continuous, surjective, and φ-equivariant, where φ acts on (0, +∞) by multiplying by λ1 . (4) The map − log Len : A → R is also continuous, surjective and φ-equivariant, where φ acts on R by adding log λ.
65
Proof. Note that (4) follows from (3). Also, in proving (3), continuity of Len follows from (2) and equivariance follows from Theorem 2.15. For surjectivity, choose a train track repe let T1 = T0 · φ, lift g to an Fr -equivariant Λ− isometry resentative g : G → G, let T0 = G, f : T0 → T1 , and interpolate f by a Stallings fold path from T0 to T1 . For each T on this f′
f ′′
path there are equivariant edge isometries T0 −→ T −−→ T1 that factor f , and so f ′′ followed by a Λ− isometry T1 → T+ is a Λ− isometry T → T+ , proving that the fold path is in b Length varies continuously and monotonically along this path from T0 to T1 with image A. [Len(G), Len(G)/λ]. Surjectivity of Len : Ab → (0, +∞) then follows from φ-equivariance. Consider a sequence Gi ∈ A which converges in Xr to some limit G. We first show that bi converges to some limit G b ∈ Xr , and second that G b ∈ A. b This the normalized sequence G b proves (1). It also proves that the normalization map A → A is continuous, but its inverse is the restriction of the continuous projection map Xbr 7→ Xr , proving (2) as well. Let T, Ti b G bi. denote the universal covers of G, Cells in Xr being locally finite, by restricting to each of the finitely many closed cells containing G we may assume that all the Gi are in a single open cell, so there is a family of marked homeomorphisms gji : Gi → Gj such that gkj ◦ gji = gki , well-defined by requiring each gji to be affine on each edge of Gi . Lift gji to an Fr -equivariant homeomorphism g˜ji : Ti → Tj which is affine on each edge. Convergence in Ab is defined by convergence of translation length functions ℓγ for all γ ∈ Fr − {Id}, whereas convergence in A is defined by convergence of ratios ℓγ /ℓγ ′ for all b i it suffices to find γ, γ ′ ∈ Fr − {Id}. Since Gi converges to G in A, to prove convergence of G a single γ ∈ Fr − {Id} such that ℓγ (Ti ) is constant. Pick any leaf L of Λ− and any i. There is a directed edge e of Gi that is traversed twice in the same direction by the realization of L in Gi . Surgery of L along e, as in the proof of Lemma 5.1, produces a legal loop in Gi , which maps to a legal loop in each Gj under gji . Choose γ ∈ Fr representing this loop. The axes of γ in Ti all correspond under the maps g˜ji , and these axes are all legal, and so b i converges to some G b ∈ Xr . ℓγ (Ti ) = ℓγ (T+ ), completing the proof that G Choose Λ− isometries fi+ : Ti → T+ . We show that, in some sense, a subsequence of fi+ b ∈ A. b converges to a Λ− isometry T 7→ T+ , proving that G Since G is in the closure of the open cell of Xr containing all of the Gi , for each i there bi → G b which collapses each component exists a unique marked homotopy equivalence pi : G of a subforest to a vertex, and which is affine on each noncollapsed edge. This subforest is trivial if and only if G is in the same open cell as all of the Gi . Note that pj ◦ gji = pi . α α b Choose a finite set A and an indexing {eα i }α∈A of the edges of Gi so that gji (ei ) = ej . b as Let B ⊂ A correspond to edges that are not collapsed by each pi . Index the edges of G β i β β b {e }β∈B so that p (ei ) = e . The edge length function of Gi , denoted ℓi : A → (0, +∞), converges to a function ℓ : A → [0, ∞) which is nonzero precisely on B. By restricting to B, b ℓ is identified with the edge length function for G. i b bi as follows. For each vertex v ∈ G, b Choose a marked homotopy equivalence q : G → G i −1 i b consider the subtree τi (v) = (p ) (v) ⊂ Gi , and choose q (v) ∈ τi (v) arbitrarily. For each b with endpoints v, w, there is a unique continuous extension of q i to eβ which edge eβ of G is affine on eβ , whose image is contained in eβi ∪ τi (v) ∪ τi (w), and such that pi ◦ q i eβ is b →G bi which is clearly a homotopic rel endpoints to the identity map. This defines q i : G 66
marked homotopy inverse for pi . Denote by p˜i , q˜i the lifts of pi , q i to Fr -equivariant maps of universal covers. b i of the trees τi (v), and so ǫi → 0 as i → ∞. Let ǫi denote the maximal diameter in G Also let ri ≥ 1 denote the maximum value of the ratios Len eβ Len eβi Len eβi + 2ǫi Len eβ , , , Len eβi Len eβ Len eβi + 2ǫi Len eβ as β varies over B, so ri → 1 as i → ∞, and ri is an upper bound for the amount by which either of the maps pi , q i stretches or compresses an edge with index in B; of course pi compresses by +∞ any edge with index in A − B. Next we show that for each vertex v of T , the points fi+ q˜i (v) vary over a compact subset of T+ . Let {E1 , . . . , EK } be the edges of the tree τi (v). For each Ek choose leaves Lk , L′k of Λ− which run through Ek and which are not asymptotic in either direction, and so the realizations of Lk and L′k in T+ intersect in a compact interval Ik independent of i. It follows that fi+ q˜i (v) ∈ I1 ∪ · · · ∪ IK , the latter being compact. Since there are only finitely many orbits of vertices we can pass to a subsequence so that fi+ q˜i (v) converges to a point denoted f (v). Extend f affinely to an Fr -equivariant map f : T → T+ . Since ri → 1 it follows that f is an edge isometry. Given a leaf L of Λ− realized i i in T , with realizations in Ti , T+ denoted Li , L+ , note that p˜ (Li ) = L, and that q˜ (L) is + Li is an isometry onto L+ , the map contained in the ǫi neighborhood of Li . Since fi fi+ q˜i L has image contained in the ǫi neighborhood of L+ , and so f (L) = L+ . The map f : L → L+ restricts to an isometry on each edge of T crossed by L, and so this map will be an isometry once we prove that it is injective. The map fi+ q˜i L can only identify a pair of points x, y ∈ L if there is a vertex v such that q˜i ([x, v] ∪ [v, y]) ⊂ τi (v), and so sup
dT (x, y) ≤ 2ri ǫi → 0
as i → ∞
x,y∈L, f i q˜i (x)=f i q˜i (y)
The map f : L → L+ is therefore an isometry. Since L is an arbitrary leaf of Λ− , we have b ∈ A. b proved that f is a Λ− isometry, and so G ♦
6.2
The Gromov topology on weak train tracks
The proof of Lemma 6.2 item (2) can be interpreted in terms of the Gromov topology on Xbr : for any D ≥ 0 and ǫ > 0, if i is sufficiently large then the restriction of q˜i to any subtree of diameter ≤ D is an ǫ-almost isometry; we verify this below. In fact, this proves one of bi converges to G b in the the easy directions of Paulin’s Theorem [Pau89], namely that if G b i converges to G b in the Gromov topology. For use in the proof of cellular topology then G Theorem 6.1, in Lemma 6.3 we present a version of this argument tailored specifically to weak train tracks, couched in terms of a relation among weak train tracks called the “Λ− relation”. This result will be applied in the verification of step 3(d) in Section 6.4. b from among the set I(T, T+ ) of all Λ− isometries T 7→ T+ Given a weak train track T ∈ A, + we pick out one, denoted kT , as follows. If Λ− is nonorientable then Theorem 5.8 says that I(T, T+ ) has a unique element, taken to be kT+ . If Λ− is orientable then, fixing an orientation, Theorem 5.8 says that I(T, T+ ) is a point or a compact oriented arc, and we take kT+ to be the positive endpoint of I(T, T+ ), characterized by T having a vertex with two or more 67
positive gates with respect to kT+ . Choosing the positive rather than the negative endpoint of I(T, T+ ) is arbitrary, the point being to make a choice which is sufficiently canonical to force the appropriate continuity property — namely, Lemma 6.3 — to be true. b the Λ− relation between T and T ′ is the set Given two weak train tracks T, T ′ ∈ A, ′ ′ R ⊂ T × T consisting of all (x, x ) with the following two properties: (1) kT+ (x) = kT+′ (x′ ) (2) There exists a leaf L of Λ− such that x is contained in the realization of L in T and x′ is contained in the realization of L in T ′ . The following lemma says that the Λ− relation can be used to exhibit closeness among weak train tracks in the Gromov topology: b ǫ > 0, any finite subset t ⊂ T , and any finite subset P ⊂ Fr , Lemma 6.3. For any T ∈ A, there exists a neighborhood U ⊂ Ab of T such that for any T ′ ∈ U , if R denotes the Λ− relation between T and T ′ , and if we define a finite subset t′ = {x′ ∈ T ′ there exists x ∈ t such that (x, x′ ) ∈ R} then the relation R ∩ (t × t′ ) is a P -equivariant ǫ-almost isometry between t and t′ .
Proof. Choose a sequence Ti converging to T . We use various notations, and we know various b i = Ti /Fr , facts, from the proof of Lemma 6.2 item (2), in particular: marked graphs G b b b b b G = T /Fr , marked homotopy equivalences pi : Gi → G, qi : G → Gi lifting to Fr equivariant maps p˜i : Ti → T , q˜i : T → Ti , etc. Sequences ri ≥ 1 converging to 1 and ǫi ≥ 0 converging to 0 were shown to have the property that for any η > 0 there exists a neighborhood U of T such that if Ti ∈ U then |ri − 1| , |ǫi | < η. It therefore suffices to prove that if ri is sufficiently close to 1 and ǫi is sufficiently close to 0 then the conclusions of the lemma hold. Consider now the Λ− isometries ki+ : Ti → T+ . We know that ki+ q˜i : T → T+ converges to a Λ− isometry fT+ : T → T+ . We wish to show that fT+ has a vertex with two or more positive gates, allowing us to conclude that fT+ = kT+ . Since there are only finitely many orbits of turns under the action of Fr , we may pass to a subsequence so that the gates in Ti with respect to the maps ki+ all correspond under the Fr -equivariant homeomorphisms g˜ji . Pick a vertex v ∈ T so that each τi (v) contains a vertex wi with two or more positive gates relative to ki+ . The construction of the maps q˜i gives us the freedom to choose q˜i (v) in τi (v), and we choose q˜i (v) = wi . It follows that there is a nondegenerate turn {e, e′ } at v such that the directions Dq˜i (e) and Dq˜i (e′ ) are contained in distinct gates of ki+ at wi . Passing to the limit, this proves that e, e′ are contained in distinct gates of fT+ at v, proving that fT+ = kT+ . Let Ri denote the Λ− relation between T and Ti : x ∈ T and y ∈ Ti are related by R if kT+ (x) = ki+ (y) and x, y are contained in realizations of the same leaf of Λ− in T, Ti , respectively. Assuming this is so, we argue that the distance between y and q˜i (x) converges to zero uniformly as i → +∞, independent of the choice of (x, y) ∈ R. Pick any leaf L of Λ− realized in T , with realization Li ⊂ Ti and L+ ⊂ T+ , so that x ∈ L and y ∈ Li . The arguments of Lemma 6.2 item (2) show that q˜i (x) is within distance ǫi of some point zi ∈ Li , and so ki+ q˜i (x) is within distance ǫi of ki+ (zi ) ∈ L+ . On the other hand ki+ q˜i (x) converges uniformly to kT+ (x) = ki+ (y) ∈ L+ , and so ki+ (zi ) converges to ki+ (y) in L+ , and thus the distance between ki+ (zi ) and ki+ (y) converges uniformly to zero. Since ki+ restricts to an 68
isometry from Li to L+ , the distance between y and zi in Li converges uniformly to zero. This proves that the distance between q˜i (x) and y converges uniformly to zero, as desired. Pick η > 0, pick a finite subset t ⊂ T , and let ti ⊂ Ti be the finite set of points y ∈ Ti related to some x ∈ t by the relation Ri . The set Ri ∩ (t × ti ) projects surjectively to t and to Ti because Ri ⊂ T × Ti projects surjectively to T and to Ti . Let D be the diameter of t. We must show that for sufficiently large i the relation Ri ∩ (t × ti ) is an η-almost isometry (equivariance with respect to any finite subset of Fr is easily verified), in other words if (x, y), (x′ , y ′ ) ∈ Ri ∩ (t × ti ) then |dT (x, x′ ) − dTi (x′ , y ′ )| < η q i (x), y) < η4 for any (x, y) ∈ Ri , it remains to show Picking i sufficiently large so that dTi (˜ η i that the restriction q˜ t is an 2 -almost isometry, that is, if x, x′ ∈ t then dT (x, x′ ) −
η η q i (x), q˜i (x′ )) ≤ dT (x, x′ ) + ≤ dTi (˜ 2 2
(6.1)
There is an integer J ≥ 0 such that for any x, x′ ∈ t the arc [x, x′ ] makes at most J turns. The restriction q˜i [x, x′ ] stretches or compresses each edge by at most ri ≥ 1, and makes a fold at each turn of length at most ǫi ≥ 0, as shown in the proof of Lemma 6.2 item 2. It follows that 1 q i (x), q˜i (x′ )) ≤ ri dT (x, x′ ) dT (x, x′ ) − Jǫi ≤ dTi (˜ ri η , from which The right hand inequality of (6.1) follows by choosing i so large that ri < 1 + 2D η ′ ′ it follows that ri dT (x, x ) ≤ dT (x, x ) + 2 . The left hand inequality follows by first choosing i so large that η2 − Jǫi > 0, and then choosing i so large that (1 − r1i )D ≤ η2 − Jǫi which implies that dT (x, x′ ) − η2 ≤ r1i dT (x, x′ ) − Jǫi . ♦
Remarks. In the course of this proof we have shown a little more regarding the strucb T+ ) = ∪ b I(T, T+ ) in the Gromov topology (we refer the reader to ture of the space I(A, T ∈A [Sko], or to Section 6.4, verification of Step 3(d), for the definition of the Gromov topology b T+ ) → A, b the on such spaces of maps). There is a continuous, open, surjective map π : I(A, b T+ ) to its domain T in A. b In “domain map”, which maps each function f : T → T+ in I(A, the above proof, we have selected a particular element kT+ from each fiber π −1 (T ) = I(T, T+ ), thereby defining a section of the map π. What we have proved amounts to the statement that b T+ ). In the case that Λ− is nonorientable this section is a continuous map k + : Ab → I(A, one has in fact that π is a homeomorphism with inverse k + . But when Λ− is orientable and when there exist nontrivial fibers, one can also select the negative endpoint, and the same b T+ ). This shows that, in some sense, the proof gives a continuous section k − : Ab → I(A, b T+ ) into intervals I(T, T+ ) all fit together in a continuous fashion, making the space I(A, b b an “interval bundle” over A whose fiber over T ∈ A is I(T, T+ ). To make sense out of this one would have to investigate local triviality, and one would have to deal with the fact that the fibers may not be topologically homogeneous, some being nondegenerate intervals and others degenerating to points. We shall not elaborate on this further, except insofar as to point out that the space of efficient representatives studied by Los and Lustig in [LL04] is b T+ ) of this interval bundle. naturally a subspace of the total space I(A, 69
6.3
Properness of the length map.
Our present goal, a piece of the proof that Len : A → (0, +∞) is a proper homotopy equivalence, is the following: Proposition 6.4 (Length is proper). The map Len : A → (0, +∞) is proper. It follows that a subset C ⊂ A is compact if and only if it is closed in A and the set Len(C) is contained in a compact subinterval of (0, +∞). The next lemma gives a compactness criterion that will be used several times in the proof of Proposition 6.4. Recall our convention from Section 2.2 that all fold maps are edge isometries. Given a subset K ⊂ Xbr define F (K; 1) to be the union of K with the set of all H ∈ Xbr for which there exists H0 ∈ K and a fold map f : H0 → H. By induction define F (K; M ) = F (F (K; M −1); 1); equivalently this is the set of all H ∈ Xbr for which there exists H0 ∈ K and a piecewise isometry f : H0 → H (not required to be an edge isometry) which factors into M marking preserving homotopy equivalences H0 7→ H1 7→ · · · 7→ HM = H each of which is either an isometry or a fold map. Lemma 6.5. For each compact K ⊂ Xbr and integer M ≥ 0 the set F (K, M ) is compact.
Proof. By induction it suffices to prove that F (K; 1) is compact. Let each closed cell σ of Xr be labelled σ(G) for some choice of marked graph G ∈ int(b σ ), so if H ∈ Xr then H ∈ int(σ(G)) if and only if there exists a marking preserving homeomorphism G ≈ H; moreover, such a homeomorphism is unique up to isotopy rel vertices. Let σ b(G) denote the preimage of σ(G) in Xbr . By local finiteness of the cell decomposition of Xr , the compact set K intersects only finitely many closed cells of Xbr , and by intersecting each of these one at a time it suffices to fix a closed cell σ b(G) and to assume that K ⊂ σ b(G). For each ordered turn (e, e′ ) of G let F (K; e, e′ ) denote the union of K with the set of all H ∈ Xbr for which there exists H0 ∈ K and a fold map f : H0 7→ H such that, with respect to the marked homeomorphism G ≈ H0 , the map f folds the turn (e, e′ ) and the length of e in H0 is greater than or equal to the length of e′ in H0 . By finiteness of the number of turns of G it suffices to fix an ordered turn (e, e′ ) of G and to prove that F (K; e, e′ ) is compact. We consider two cases, depending on the valence of the common initial vertex v of the edges e, e′ . Suppose first that v has valence ≥ 4. Let G# be the marked graph obtained from G by folding proper initial segments of e and of e′ to obtain an edge e# of G# . The cell σ b(G) is a codimension 1 face of σ b(G# ), defined by a marked homotopy equivalence G# → G that collapses e# to a point and is otherwise a bijection. Under this collapse map, the edges e, e ′ of G correspond to two edges of G# also denoted e, e′ . Let ∆ = {(H0 , t) ∈ σ b(G) × [0, ∞) # ′ b(G ) taking (H0 , t) to the element 0 ≤ t ≤ LenH0 (e ) ≤ LenH0 (e)}. Define a map ∆ 7→ σ of σ b(G# ) which assigns length t to e# , which subtracts t from the lengths of the edges e, e′ in H0 , and which leaves the lengths of all other edges of H0 unchanged. Notice that when t = 0 the length assigned to e# equals zero; whereas when t = LenH0 (e′ ) then the length assigned to e′ equals zero and, in the case LenH0 (e′ ) = LenH0 (e), the length assigned to e also equals zero; and so in either of these two cases the image of (H0 , t) is in a face of the ′ closed cell σ b(G # ). The map ∆ 7→ σ b (G# ) is clearly continuous, and F (K; e, e ) is the image under this map of a compact subset of ∆, namely the set {(H0 , t) ∈ ∆ H0 ∈ K}, proving that F (K; e, e′ ) is compact. 70
Suppose next that v has valence 3. Let e′′ be the third edge incident to v. Note that e′′ is distinct from both e and e′ , for if say e′′ = e, consider any oriented leaf L of Λ− travelling backwards along e; when L hits the vertex v, it cannot take the illegal turn (e, e′ ) and so L must continue into e′′ = e again, and this must repeat forever, contradicting surjectivity of half-leaves of Λ− . Now define ∆ as above, and define the map ∆ 7→ σ b(G) taking (H0 , t) to the element of σ b (G) which subtracts t from the lengths of e, e′ and adds t to the length of e′′ . The proof is then completed as in the previous case. ♦
Proof of Proposition 6.4. Since Ab is a closed subset of Xbr and since Len is continuous, for −1 b each closed interval [a, b] ⊂ (0, +∞) the set Ab[a,b] = A∩Len [a, b] is closed, and it suffices to b prove that A[a,b] is compact. Lemma 6.6 below implies that there exists a train track G ∈ Ab and an integer M ≥ 0 such that for each H ∈ Ab[a,b] there is a Λ− isometry G 7→ H which has a Stallings fold sequence consisting of at most M folds. It follows that Ab[a,b] ⊂ F ({G}; M ), and the latter set is compact by Lemma 6.5. Since Ab[a,b] is a closed subset of a compact set, it is compact. This completes the proof subject to our verifying Lemma 6.6. ♦ Lemma 6.6. For any closed interval [a, b] ⊂ (0, +∞) there exists a train track map g : G → b and there exists an integer M ≥ 0, so that if H ∈ Ab[a,b] , G with G normalized to be in A, e → T+ is a Λ− -isometry, then there exists an edge isometry h : G → H such and if fHe : H ˜ and for any such edge isometry h : G → H, the combinatorial length of h is that fg = fHe h, at most M , and h has a Stallings fold sequence of length at most M . Proof. Pick a train track G ∈ A with the finest possible Nielsen structure. Pick ǫ > 0 for G by applying Proposition 5.4. We may assume that ǫ > b, for otherwise we replace G by φ−n (G) for sufficiently large n, which has the effect of replacing ǫ by ǫ · λ(φ)n . By e → T+ there is an edge Proposition 5.4, for every H ∈ A[a,b] and every Λ− isometry fHe : H ˜ where h ˜ is a lift of h : G → H. isometry h : G → H so that fg = fHe h ˜ Recall from the discussion Let h : G → H be any edge isometry such that fg = fHe h. preceding Lemma 2.6 that, after subdividing H at the h-images of the vertices of G, the subdivided H has at most 5r − 5 edges. Since Len(H) ≥ a, there exists an edge e of H with Len(e) ≥ a/(5r − 5). For any edge isometry G′ → H whose vertices map disjointly from e, define an e-arc in G′ to be an open arc of G′ which maps isometrically onto the interior of e. Two e-arcs in G′ are disjoint and each e-arc has length equal to Len(e), and so G′ has at most Len(G′ )(5r − 5)/a distinct e-arcs. Using the notation of (2.2), apply Lemma 2.6 to obtain a Stallings fold sequence G = g2 g1 gJ G0 −→ G1 −→ · · · −→ GJ = H for h whose length J is bounded by the combinatorial length of h. For j = 0, . . . , J let Mj be the number of e-arcs in Gj . The sequence Mj is nonincreasing: each e-arc in Gj−1 maps isometrically onto an e-arc in Gj ; the e-arcs in Gj−1 are precisely the components of the inverse images of the e-arcs of Gj ; and so either Mj = Mj−1 , or Mj < Mj−1 which occurs precisely when the fold map gj : Gj−1 → Gj identifies at least one pair of e-arcs in Gj−1 to a single e-arc of Gj . There is therefore a sequence 0 ≤ j1 < · · · < jN = J of some length N such that the subsequence Mj , for jn−1 < j ≤ jn , is constant, and such that Mjn > Mjn +1 . It follows that N ≤ M0 ≤ Len(G)(5r − 5)/a. The following lemma, applied with K equal to the singleton {G}, and with h equal to the composed map G0 7→ Gj1 −1 , implies that Gj1 −1 belongs to a compact subset K ′ 71
of Xbr and that the map G0 7→ Gj1 −1 has uniformly bounded combinatorial length. Then apply Lemma 6.5 to the single fold Gj1 −1 7→ Gj1 to obtain a compact set K ′′ that contains Gj1 . Then reapply the following lemma with K = K ′′ and with h equal to the composed map Gj1 7→ Gj2 −1 . Continue inductively in this manner to prove that each composed map Gjn−1 7→ Gjn −1 has uniformly bounded combinatorial length. Summing up we obtain a uniform bound M for the combinatorial length of h. Applying Lemma 2.6, h has a Stallings fold sequence of length at most M . ♦ Lemma 6.7. For any compact set K ⊂ Xbr there exists a compact set K ′ ⊂ Xbr and an integer L ≥ 0 so that for any G ∈ K, G′ ∈ Xbr , and edge isometry h : G → G′ , if ℓγ (G′ ) ≥ ℓγ (T + ) for all γ ∈ Fr , and if there exists an edge of G′ over the interior of which h is injective, then G′ ∈ K ′ and h has combinatorial length ≤ L. In this lemma, edges of G′ are determined as usual by subdividing at the images of the vertices of G under the map h, and so G′ has at most 5r − 5 edges. Proof. A proper, connected, nontree subgraph of a marked graph G determines a conjugacy class of proper, nontrivial free factors of Fr by pulling back the fundamental group of the subgraph under the marking Rr → G. Let F F(K) be the set of all free factor conjugacy classes associated to the proper, connected, nontree subgraphs of G, as G varies over K. This set is finite, because the cells of the cellular topology on Xbr are locally finite (see Section 2.1). By Theorem 5.4 of [BFH97], for each proper free factor conjugacy class A in Fr , the infimum of ℓγ (T + ) for γ ∈ A is a positive number denoted ǫ(A). Let ǫ > 0 be the minimum value of ǫ(A) as A varies over the finitely many elements of F F(K). If G, G′ are as in the lemma then it follows that for each γ ∈ Fr , if the circuit in G representing γ does not surject onto G, then the circuit in G′ representing γ has length ≥ ǫ. Let e′ ⊂ G′ be an edge over whose interior h is injective, let X ′ be the complement in G′ of the interior of e′ , and let X = h−1 (X ′ ) ⊂ G. Then h(X) = X ′ and, since h is a homotopy equivalence, h restricts to a homotopy equivalence between X and X ′ . It follows that X, X ′ have the same number of components — two or one depending on whether e′ is separating or nonseparating — these components are connected, proper, subgraphs of G, G′ , respectively, and neither X nor X ′ is a tree for otherwise G or G′ would have a valence 1 vertex. Moreover, components of X, X ′ that correspond under h represent the same proper, nontrivial free factor conjugacy class of Fr . Thus every simple circuit in X ′ has length at least ǫ and so crosses an edge of length at least ǫ/(5r − 6), the subgraph X ′ having at most 5r − 6 edges. Letting E ′ be the union of e′ and the set of edges of G′ with length at least ǫ/(5r − 6), it follows that every simple circuit in G′ intersects an edge in E ′ , and so the complement of E ′ is a forest U ′ . For each edge e of G, since Len(e) ≤ Len(G), the edge path h(e) contains at most 1 + Len(G) · (5r − 6)/ǫ elements of E ′ . Removing these edges from the edge path h(e) leaves at most 2 + Len(G) · (5r − 6)/ǫ complementary subpaths each of which is contained in the forest U ′ and so is embedded, crossing each edge of U ′ at most once. We conclude that there is a uniform bound to the combinatorial length of the edge path h(e). Since e was arbitrary, the combinatorial length of h is uniformly bounded. The map h therefore has a Stallings fold sequence of uniformly bounded length. Applying Lemma 6.5, G′ lies in a compact subset K ′ of Xr . ♦
72
6.4
Applying Skora’s method to the Properness Theorem 6.1
Let A ⊂ Xr be the set of weak train tracks of a nongeometric, fully irreducible φ ∈ Out(Fr ), with a normalization A ֒→ Xbr determined as usual by a choice of normalization of Tφ+ . Rather than continue with the notation Ab for the image of the normalization map, in this section we assume implicitly that all marked graphs and Fr -trees representing elements of A are normalized. The major tool that we use in proving Theorem 6.1, that the map Len : A 7→ (0, +∞) is a proper homotopy equivalence, is Skora’s method [Sko] for detecting the homotopy type of various spaces of trees. The preprint [Sko] being unpublished and therefore difficult to obtain, other accounts to consult include [Whi93], [Cla05], and [GL05]; we shall quote a technical result from the latter which simplifies the method. While Skora applies his technique to very general spaces of group actions on trees, we shall consider only free, simplicial, minimal Fr -trees, that is, elements of outer space Xr . Spaces of edge isometries. Skora’s method requires defining topologies on various spaces of piecewise isometries between trees. We adopt here an addition to the method suggested by Definition 3.8 and Proposition 3.10 of [GL05] which guarantees that minimality of trees is true at all stages of Skora’s method, and which allows us to focus solely on edge isometries among free, simplicial, minimal Fr trees, that is, elements of Xbr . Given free, simplicial, minimal Fr trees S, T , a piecewise isometry f : S → T satisfies the minimality condition if for each x ∈ S there exists an open arc embedded in S and containing x such that the restriction of f to this arc is an isometry; equivalently, f is an edge isometry and each vertex of S has at least two gates with respect to S. The motivation for the minimality condition is that if it failed, if there were a point x having a single gate, then one could construct an equivariant Stallings fold sequence for f by first folding all of the turns at x, producing a tree which has a valence 1 vertex and therefore is not minimal. Particularly important is the fact that Λ− isometries automatically satisfy the minimality condition. Let M be the set of equivalence classes of piecewise isometries f : S → T satisfying the minimality condition, where S, T are free, simplicial, minimal Fr trees, and where f : S → T and f ′ : S ′ → T ′ are equivalent if there exist Fr equivariant isometries h : S → S ′ , k : T → T ′ such that kf = f ′ h. Formally the element f : S → T is to be regarded as an ordered triple (f, S, T ) which allows us to define two projections Dom, Ran : M → Xr , the domain projection Dom(f, S, T ) = S and the range projection Ran(f, S, T ) 7→ T . The group Out(Fr ) acts on M from the right so that the domain and range projections are Out(Fr ) equivariant. The action is defined as follows. Given f : S → T in M and given φ ∈ Out(Fr ), choose Φ ∈ Aut(Fr ) representing φ. Recall that the Fr -trees S · φ, T · φ have underlying R-trees S, T but with Fr actions precomposed by Φ. The Fr -equivariant map f : S → T therefore remains Fr -equivariant when the actions are precomposed by Φ, and we define this map to be the image of f under φ, well-defined in M independent of the choice of Φ. We put off until later the definition of the Gromov topology on M, as given in the references above. For the moment, we note that the domain and range projections Dom, Ran : M → Xr are continuous maps with respect to this topology.
73
1 1 Notice that the reparameterized maps − log Len = log Len :A → : A → R and Len i (0, +∞) are “correctly oriented” in the sense that, on an orbit Gi = φ (G0 ), the values of these maps increase as i increases. The map Len : A → (0, +∞), on the other hand, is “incorrectly oriented”, and the map − log Len carries heavier notational baggage, so in the 1 . proof that follows we often think in terms of the map Len Proof of the Properness Theorem 6.1. We give the proof in the several steps. 1 Step 1: A bi-infinite fold path L. Our goal being to prove that Len : A → (0, +∞) 1 :L→ is a proper homotopy equivalence, we begin by choosing a subset L ⊂ A such that Len (0, +∞) is a homeomorphism. This subset L will be a particular fold line, periodic under an appropriate power of φ. Passing to a positive power of φ, we may assume φ is rotationless. Applying Proposition 4.4, choose g : G → G to be a train track representative of φ whose local Whitehead decomposition is as fine as possible, meaning that the local stable Whitehead graphs of g have no cut points. We assume that the train track G is normalized just as A is, and so e → T+ . Define G0 = G and Gi = λ−i φi (G) for there is an associated Λ− isometry fg : G all i ∈ Z. The map g : G → G induces marked edge isometries gi+1,i : Gi → Gi+1 . Let Li = Len(Gi ) = λ−i Len(G0 ). For later use recall that if we define gji = gj,j−1 ◦ gj−1,j−2 ◦ · · · ◦ gi+1,i : Gi → Gj then the e i and the Fr -equivariant edge isometries g˜ji form a direct system universal covering trees G in the category of metric spaces and distance nonincreasing maps, the direct limit of this e i → T+ such that the two maps system is the Fr tree T+ , we obtain direct limit maps g˜i+ : G + e e fg , g˜0 : G = G0 → T+ are equal to each other. As described in Section 2.2, choose a Stallings fold sequence for g1,0 and interpolate the fold maps to get a Stallings fold path for g1,0 connecting G0 to G1 . We parameterize this 1 , with parameter interval [ L10 , L11 = Lλ0 ]. Translating this path by φi and fold path by Len reparameterizing, one obtains a Stallings fold path for gi−1,i connecting Gi−1 to Gi and i i+1 1 , with parameter interval [ Lλ0 , λL0 ]. Concatenating these paths over parameterized by Len all i ∈ Z we obtain a bi-infinite Stallings fold path (0, +∞) 7→ A. The image of this path is the desired set L, completing Step 1. 1 Note that Len : L → (0, +∞) is a homeomorphism. The inverse of this homeomorphism, which we denote ρ : (0, +∞) → L, will be our preferred parameterization for the fold path L.
Step 2: The map A → L. We construct a proper, continuous map σ : A → L whose restriction to L, while not the identity on L, is properly homotopic to the identity on L. In later steps we shall prove that σ is a proper homotopy equivalence. Applying Proposition 5.4, there exists a number η > 0 such that if T is any weak train track with Len(T ) < η and if fT : T → T+ is a Λ− isometry, then there exists a Λ− isometry e 0 → T such that fT ◦ ψ = fg . ψ: G Define the map σ : A → L by η −1 η =ρ ρ (T ) σ(T ) = ρ L0 λ Len(T ) L0 λ 74
This map is evidently continuous, and it is proper by Proposition 6.4. The restriction to L, when conjugated by ρ, is the map from (0, +∞) to itself given by r 7→ Lη0 λ r. This map being an orientation preserving homeomorphism of (0, +∞), the map σ L : L → L is properly homotopic to the identity. η < 1. In other Remark. For later purposes we note that η < L0 < λL0 and so λL 0 1 words, the map σ decreases the value of the parameter Len , multiplying it by the number η λL0 , and so σ moves each point of L strictly to the left.
Step 3: The map to M. For each T ∈ A we define a Λ− isometry ψ T : σ(T ) → T with the property that the map (σ(T ), T ) 7→ ψ T is a continuous map from the set {(σ(T ), T ) ∈ L × A T ∈ A} to the set M. Letting u ∈ (0, +∞) be such that σ(T ) = ρ(u), the map ψ T : σ(T ) → T will be defined by factoring a particular Λ− isometry fu+ : ρ(u) = σ(T ) → T+ through T , where the other factor is a particular Λ− isometry kT+ : T → T+ . We construct these Λ− isometries, and we define ψ T and verify its properties, in several substeps. Step 3(a): The Λ− isometry fu+ from ρ(u) ∈ L to T+ . Define the semiflow of the flow line L as follows. This semiflow will be a family of Fr -equivariant Λ− isometries fts : ρ(t) → ρ(s), 0 < s < t < +∞, so that the semiflow identity fut ◦ fts = fus is satisfied. For each integer i the map f1/Li ,1/Li−1 : ρ(1/Li−1 ) → ρ(1/Li ) is defined to be the map ei−1 → G ei . For each integer i, the subfamily of maps fts for 1 ≤ s ≤ t ≤ 1 is g˜i,i−1 : G Li−1 Li (up to reparameterization) defined to be the semiflow of the chosen Stallings fold sequence for gi,i−1 : Gi → Gi−1 . The definition of fts extends uniquely to 0 < s ≤ t < ∞ in such a way as to satisfy the semiflow identity. The trees ρ(t) and the maps fts form a direct system in the category of Fr -equivariant metric spaces and distance nonincreasing maps. The direct limit of this direct system is the e i and the maps gji form a co-finite subsystem, whose direct Fr tree T+ , because the trees G limit is known to be T+ by Theorem 2.13. We obtain direct limit maps fu+ : ρ(u) → T+ which satisfy the property that fu+ ◦ fut = ft+ for 0 < t < u < +∞. When ρ(u) = σ(T ) we + also denote fu+ = fσ(T ) : σ(T ) → T+ . Step 3(b): The Λ− isometry kT+ from T ∈ A to T+ . This was defined in Section 6.2 using Theorem 5.8: if Λ− is nonorientable then kT+ is the unique Λ− isometry T 7→ T+ ; and when Λ− is oriented then kT+ is the positive endpoint of the compact oriented interval I(T, T+ ) of Λ− isometries. Step 3(c): The Λ− isometry ψ T : σ(T ) → T . For each T ∈ A the map ψ T is + characterized as the unique Λ− isometry from σ(T ) to T with the property that fσ(T ) = + T T kT ◦ ψ . Uniqueness of ψ follows from the proof of uniqueness in Proposition 5.4. However, we still have to establish that ψ T exists with this property. The existence of ψ T being invariant under the action of φ on A, it suffices to prove existence after replacing T by φk (T ) for a particular power k. Since φ increases the parameter
75
1 Len
by multiplying it by the quantity λ, after acting by the appropriate power of φ, we may assume that 1 1 λ ≤ ≤ η Len(T ) η Recalling that
1 Len(σ(T ))
=
η L0 λ Len(T )
we have
1 1 1 ≤ ≤ L0 λ Len(σ(T )) L0 e is therefore defined. Since : σ(T ) → G e = ρ(1/L0 ) → T with the Len(T ) ≤ η, Proposition 5.4 gives us a Λ− isometry hT : G + T property that fg = kT ◦ hT . Define ψ = hT ◦ f L1 ,t0 , and so
Setting t0 =
1 Len(σ(T )) ,
the semiflow map f L1
0
,t0
0
kT+
T
◦ψ =
kT+
◦ hT ◦ f
= fg ◦ f L1
0
=f =
+
◦f
1 L0
ft+0
=
1 L0
,t0
,t0
1 L0
,t0
+ fσ(T )
as desired. Remark: What we have done is to define ψ T first for T in a fundamental domain of the action of φ on A, and then to extend over all of A by φ-equivariance. This raises the issue 1 1 of well-definedness of ψ T where one end of the fundamental domain, defined by Len(T ) = η, λ 1 is translated by φ to the other end of the fundamental domain, defined by Len(T ) = η . This T issue is made moot by the fact that ψ is already uniquely characterized: it is the unique Λ− isometry from σ(T ) → T that satisfies the equation ft+ = kT+ ◦ ψ T . Step 3(d): Continuity of the map (σ(T ), T ) 7→ ψ T . We must prove that the map from the set {(σ(T ), T ) T ∈ A} ⊂ L × A to the set M, taking the pair (σ(T ), T ) to the map ψ T : σ(T ) → T , is a continuous map. This is the most technical of the steps and requires us to look carefully at the topology of the mapping space M. We put this issue off for a while more, in order to get through the steps in the proof of Theorem 6.1. Step 4: Combining Steps 2 and 3, we obtain a continuous map A 7→ M defined by T 7→ ψ T , so that postcomposing with Dom : M → Xr we obtain the map σ : A 7→ L, and postcomposing with the map Ran : M → Xr we obtain the identity map on A. Now apply the following theorem: Theorem 6.8 ([Sko], [GL05]). Letting ∇ = {(s, t) ∈ [0, ∞] × [0, ∞] s ≤ t}, there exists a continuous map θ : M × ∇ → M, denoted formally as θ(ψ, s, t) = (ψts , Ts , Tt ) with the following properties for each ψ : S → T in M: 76
(1) The map ψ∞,0 : T0 → T∞ equals ψ : S → T . (2) The map ψ0,0 : T0 → T0 equals the identity on S. (3) The map ψ∞,∞ equals the identity on T . (4) (The semiflow property) ψut ◦ ψts = ψus for 0 ≤ s ≤ t ≤ u ≤ ∞. Remarks on Theorem 6.8. Rather than the general Fr trees allowed in [Sko], we are restricting elements (ψ, S, T ) ∈ M by requiring S, T to be free, simplicial, minimal Fr trees, and by requiring f to satisfy the minimality condition. To see that this is allowable, first note that if M is defined using ordinary piecewise isometries rather than piecewise isometries satisfying the minimality condition, then Theorem 6.8 is exactly the main result, Theorem 4.8, of [Sko] except for the restrictions to free simplicial trees (and except for correcting the erroneous {(s, t) ∈ [0, 1] × [0, 1]} occurring in the statement of Theorem 4.8). Then note that for a map θ(ψ, s, t) = ψts : Ts → Tt obtained by the constructions of [Sko] or of [GL05], the tree Tt is simplicial by Proposition 3.6 of [GL05], and Tt is free because T is free, and ψts satisfies the minimality condition by Propositions 3.4, 3.10, and 3.11 of [GL05]. To complete Step 4, as a consequence of Theorem 6.8 the following chain of compositions is continuous: (T,s,t)7→(ψ T ,s,t)
(T,s)7→(T,s,∞)
θ
Dom
A × [0, ∞] −−−−−−−−−−→ A × ∇ −−−−−−−−−−−→ M × ∇ − → M −−−→ Xr This homotopy, which we denote h : A × [0, ∞] → Xr , has the property that h(T, 0) = σ(T ) and h(T, ∞) = T , in other words h is a homotopy from the map σ : A 7→ L to the identity map on A. As a consequence of the semiflow property, for each t ∈ [0, ∞] we obtain a commutative diagram as follows, where we denote Tt = h(T, t): ψT
σ(T ) = T0 II II II II T II ψt,0 $
Tt
/ T∞ = T ; ww ww w ww T ww ψ∞,t
+ kT
/ T+
T T We reiterate that the maps ψt,0 , ψ∞,t ∈ M are edge isometries with the minimality property.
Step 5: We must verify that h takes values in A, in other words, in the above diagram the tree Tt is a weak train track. For any leaf L of Λ− , its realization in σ(T ) denoted Lσ(T ) T maps injectively under ψ T and so also maps injectively under the edge isometry ψt,0 , with T image therefore equal to the realization of L in Tt , denoted Lt . Note that ψ∞,t also restricts to an injection on Lt , and its image equals ψ T (Lσ(T ) ), which is the realization of L in T T T denoted LT . Since ψ∞,t is an edge isometry it follows that ψ∞,t restricts to an isometry + T from Lt to LT . The map kT ◦ ψ∞,t therefore restricts to an isometry from Lt into T+ . This T being true for any leaf L of Λ− , we have proved that kT+ ◦ ψ∞,t : Tt → T+ is a Λ− isometry, and so Tt ∈ A as desired. 77
Step 6: We must verify that the homotopy h : A × [0, ∞] → A is a proper map, that is, for each compact set C ⊂ A the set h−1 (C) is compact. This will be a consequence of 1 Theorem 6.4. Since C is compact, there is a closed interval [a, b] ⊂ (0, ∞) such that Len C takes values in [a, b]. For each T ∈ A, the definition of h and the remark at the end of the 1 verification of Step 2 combine to show that as t goes from 0 to ∞ the value of Len(h(T,t)) η 1 1 1 1 increases monotonically from Len(h(T,0)) = Len(σ(T )) = λL0 Len(T ) to Len(h(T,∞)) = Len(T ) . It follows that η 1 a, b ⊂ Len(h−1 (C)) λL0 1 that is, the function Len restricted to h−1 (C) takes values in a closed subinterval of (0, +∞). −1 Since h is continuous, h (C) is also closed. Applying Proposition 6.4, it follows that h−1 (C) is compact.
Steps 1 through 6 prove that the map σ : A → L is a proper homotopy equivalence. Composing this map with the homeomorphism Len : L → (0, +∞), and then composing with 0 the self homeomorphism of (0, +∞) which is multiplication by the number λL η , we obtain the map Len : A → (0, +∞), proving that this map is a proper homotopy equivalence. This finishes the proof of Theorem 6.1, subject to completing the verification of Step 3(d). T Verifying Step 3(d): Continuity of the map (σ(T ), T ) 7→ ψ . Let P = {(σ(T ), T ) T ∈ A}, so P equals the subset of L × A consisting of all (S, T ) such η η that Len(T )/ Len(S) = λL , equivalently, all (ρ(u), T ) such that u Len(T ) = λL . We shall 0 0 show that the map P 7→ M taking (ρ(u), T ) to ψ T is continuous. We remind the reader that ψ T : ρ(u) → T is the unique Λ− isometry that factors the direct limit Λ− isometry + fρ(u) : ρ(u) → T+ through the Λ− isometry kT+ : T → T+ . In this argument we shall denote ψ T = ψuT , the superscript being the range and the subscript denoting the domain ρ(u). Recall from Section 2.1 the Gromov topology on Xr , with basis element V (T ; ǫ, t, P ) determined by T ∈ Xbr , ǫ > 0, a finite subset t ⊂ T , and a finite subset P ⊂ Fr , consisting of all T ′ ∈ Xbr for which there exists a finite subset t′ ⊂ T ′ and a P -equivariant ǫ-almost isometry ρ ⊂ t × t′ that projects surjectively to t and to t′ . Next we review from [Sko] the Gromov topology on M. A basis element U (ψ; ǫ, s, t, P ) for this topology is determined for each ψ : S → T in M, ǫ > 0, finite subsets s ⊂ S, t ⊂ T , and finite set P ⊂ Fr . This basis element consists of all ψ ′ : S ′ → T ′ ∈ M for which there exist finite subsets s′ ⊂ S ′ and t′ ⊂ T ′ , and P -equivariant ǫ-almost isometries R0 ⊂ s × s′ , R1 ⊂ t × t′ such that (x, y) ∈ R0 implies (ψ(x), ψ ′ (y)) ∈ R1 . To put this more colloquially, ψ ′ : S ′ → T ′ is a member of U (ψ; ǫ, s, t, P ) if we can exhibit S ′ as a member of V (S; ǫ, s, P ) and T ′ as a member of V (T ; ǫ, t, P ) in such a way that the two P -equivariant ǫ-almost isometries that exhibit these memberships are respected by the maps ψ and ψ ′ . Note the evident continuity of the domain and range maps Dom, Ran: M → Xbr . We remark that the Gromov topology on M as defined in [Sko] uses arbitrary compact subsets of metric spaces, not just finite subsets. However, just as Paulin does in Proposition 4.1 of [Pau89], one obtains a different basis for the same topology on M by restricting to finite subsets. The relations that we shall use for ǫ-almost isometries will be restrictions of the Λ− relations defined in Section 6.2.
78
Pick (ρ(u), T ) ∈ P, denote S = ρ(u), and let ψ = ψuT : S → T . Pick a basis element U (ψ; ǫ, s, t, P ) ⊂ M about ψ. Corresponding to this are basis elements V (S; ǫ, s, P ) ⊂ Xr about S and V (T ; ǫ, t, P ) ⊂ Xbr about T . We must produce P -equivariant ǫ-almost isometries which exhibit membership in these two basis elements. Applying Lemma 6.3, we obtain a neighborhood W ⊂ A of T so that for each T ′ ∈ W , there is a finite subset t′ ⊂ T ′ such that if R1 denotes the restriction to t × t′ of the Λ− relation between T and T ′ , then R1 surjects to t and to t′ , and R1 is a P -equivariant ǫ-almost isometry, thereby exhibiting that T ′ ∈ V (T ; ǫ, t, P ). We also need a relation R0 that exhibits membership in V (S; ǫ, s, P ) as long as |u − u′ | is sufficiently small. We will define this relation in two cases, depending on whether u ≤ u′ or u′ < u. When u ≤ u′ then the semiflow associated to the Stallings fold path L gives a Λ− isometry fu′ ,u : S = ρ(u) → ρ(u′ ) = S ′ . Set s′ = fu′ ,u (s). Regarding fu′ ,u as a subset of S × S ′ , let R0 = fu′ ,u ∩ (s × s′ ); to put it another way, R0 is the restriction of fu′ ,u to s. P -equivariance of R0 is an immediate consequence of Fr -equivariance of fu′ ,u . There is an integer J such that any two points of the finite set s are connected by an edge path with no more than J turns. We choose |u − u′ | sufficiently small so that the map fu′ ,u is a fold of length no more than ǫ/J. So, if x, y ∈ s then the segment between x and y undergoes at most J folds each of length at most ǫ, and so the distance dS ′ (fu′ ,u (x), fu′ ,u (y)) is less than dS (x, y) by an amount that is no more than J(ǫ/J) = ǫ. This shows that fu′ ,u s × s′ is an ǫ-almost isometry. Next assume that u′ ≤ u. The semiflow gives Λ− isometry fu,u′ : S ′ = ρ(u′ ) → S = ′a −1 −1 ′ ρ(u). Let s = fu,u′ (s) and let R0 = (fu,u′ s ) ⊂ s × s′ . P -equivariance is again immediate. Letting H(s) be the convex hull of s in S, we can think of s′ as obtained from s by splittings at certain vertices of H(s). To be precise, for any number η > 0 such that η is less than half the length of the shortest edge of H(s), if S ′ is sufficiently close to S in Xbr then s′ is obtained from s in the following manner: for certain vertices v in H(s), choose a direction d at v, choose a segment [v, w] representing d of length at most η, partition the directions at v distinct from d into two nonempty subsets D1 ∪ D2 , and then split the segment [v, w] into two segments [v1 , w] and [v2 , w], reattach the directions D1 to v1 , and reattach the directions D2 to v2 . From this construction it follows that there is a constant J depending only on s such that any two points x, y ∈ s′ are connected by an edge path with no more than J turns. This implies that dS (fu′ ,u (x), fu′ ,u (y)) is less than dS ′ (x, y) by an −1 s × s′ is amount that is no more than Jη, and so by choosing η < ǫ/J it follows that fu,u ′ an ǫ-almost isometry. ′ Finally, we must verify that R0 and R1 are respected by the maps ψuT and ψuT′ . That ′ is, we must show that if (x, y) ∈ R0 then (ψuT (x), ψuT′ (y)) ∈ R1 . Focussing on the case that u ≤ u′ (the opposite case u′ ≤ u being handled similarly), we have y = fu′ ,u (x), and we must verify that ′ ′ (ψuT (x), ψuT′ ◦ fu′ ,u (x)) = (ψuT (x), ψuT (x)) ∈ R1 (6.2) R1 being the restriction of the Λ− isometry between T and T ′ , there are two things to check. ′ First, if L is any leaf of Λ− realized in ρ(u) such that x ∈ L, then since ψuT and ψuT are ′ both Λ− isometries it follows that ψuT (x) is in the realization of L in T , and ψuT (x) is in the realization of L in T ′ . Second, we note that ′
kT+ ◦ ψuT (x) = fu+ (x) = kT+′ ◦ ψuT (x) 79
completing the verification of (6.2), and thereby completing the verification of Step 3(d) and the proof of Theorem 6.1.
80
7
Fold lines
In this section we investigate Definition (1) of the axis bundle of a nongeometric, fully irreducible φ ∈ Out(Fr ): the union of fold lines converging in X r to T−φ in the negative direction and to T+φ in the positive. Our main result, Theorem 7.2, says that every point on such a fold line is a weak train track for φ, thus proving that Definition (1) of the axis bundle implies Definition (2). We also prove the converse, that Definition (2) implies Definition (1); see Proposition 7.1. Recall from Section 2.3 that a fold path in Xr is an injective, continuous, proper map γ : I → Xr with domain a connected subset I ⊂ R, for which there is a continuous lift γ : I → Xbr , marked graphs Gr representing γ b b(r) for each r ∈ I, and a family of edge isometries fsr : Gr → Gs defined for s ≤ r in I and satisfying the semiflow condition fts ◦ fsr = ftr . Reparameterizing γ if necessary, we ofttimes impose the following: Normalization Condition: Len(Gr ) = e−r for each r ∈ I. If γ is normalized then Gr is uniquely determined up to an isometry that preserves marking, and so the lift γ b is uniquely determined. A translation in the parameter amounts to a change in normalization: for each T ∈ R, the reparameterization γ ′ (r′ ) = γ(r′ + T ) is also a fold path, and γ b′ (r′ ) = eT · γ b(r′ + T ). The set of data (Gr )r∈I , (gsr : Gr → Gs )r<s∈I is called a connection for the fold line γ. While (Gr ) is uniquely determined by γ in the presence of the normalization condition, the maps (gsr ) are not determined in general, as we saw from the example in Section 5.5. A normalized fold path γ whose domain I is noncompact only on the positive end then γ is a fold ray; and if I is noncompact only on the negative end then γ is called a split ray. Denote I− = inf(I) ∈ R ∪ {−∞}, I+ = sup(I) ∈ R ∪ {+∞} As we shall see in the arguments below, if I is noncompact on the negative end then I− = −∞, in other words Len(Gr ) diverges to +∞ as r → I− . But we shall shortly give an example in which I is noncompact on the positive end, γ is indeed proper on the positive end as required, and yet I+ 6= +∞, that is, Len(Gr ) may converge to a positive number as r → I+ .
7.1
Examples of fold paths
In this section we give several examples and constructions of fold paths. For each nongeometric, fully irreducible φ ∈ Out(Fr ) we construct a periodic fold line through every train track for φ, and we construct a fold line for φ passing through every weak train track for φ; the latter construction proves that Definition (1) of the axis bundle of φ implies Definition (2). In order to motivate some of the later results, we shall also describe a fold ray whose length does not converge to zero, and a split ray which does not converge to a unique point in compactified outer space. Example: periodic fold lines. Given a nongeometric, fully irreducible φ ∈ Out(Fr ) with expansion factor λ = λ(φ), we outlined in the introduction the construction of a φperiodic fold line. To give the construction in detail, consider a train track map g : G → G representing φ. Normalize G ∈ Xbr by requiring Len(G) = 1. Let Gi = λ1i G · φ ∈ Xbr . 81
For integers i ≤ j let gji : Gi → Gj be the marked homotopy equivalence which agrees with g j−i : G → G. Following the method described in Section 2.3, choose a Stallings fold sequence for g1,0 : G0 → G1 , choose an interpolating path for each fold map in that sequence, and concatenate these paths to get a Stallings fold path γ b0 connecting G0 to G1 . The path 0 0 0 γ , when normalized, becomes γ b b : [0, log λ] → Xbr with γ b (0) = G0 and γ b0 (log λ) = G1 . For each integer i define the path γ bi : [i log λ, (i + 1) log λ] → Xbr to be γ bi (t) = (b γ 0 · φi )(t − i log λ). i b Define the path γ b : R → Xr to be the concatenation of γ b over all integers i. Connection data for γ b is easily produced starting from the maps gji and connection data for γ b0 . Define the path γ : R → Xr by projectivizing γ b. To prove that γ is proper, it suffices to prove that γ(t) converges in X r to T−φ as t → −∞ and to T+φ as t → +∞: by construction the path γ satisfies the φ-periodicity condition γ(t − i log(λ)) = γ(t) · φi , and the desired convergence properties follow from the uniform source-sink dynamics of the action of φ on compact subsets of Xr [BFH97]. Example: Weak train tracks. We prove here that Definition (2) of the axis bundle implies Definition (1), by constructing a fold line through every weak train track: Proposition 7.1. Let φ ∈ Out(Fr ) be fully irreducible and let Aφ ⊂ Xr denote the set of weak train tracks of φ. For each H ∈ Aφ there is a fold line γ : R → Xr passing through H such that γ(t) → T± = T±φ as t → ±∞. Proof. Choose a normalization of T+ . Let Abφ ⊂ Xbr be the corresponding normalization of Aφ , as discussed in the beginning of Section 6. Denote the normalizations of H and f by e 0 → T+ . H0 ∈ Abφ and by f0 : H ˜1 : H e0 → H e1 As in the proof of Lemma 2.6, let h1 : H0 → H1 be a fold map whose lift h ˜ e factors the map f0 with induced map f1 : H1 → T+ , and such that (h1 and) h1 are complete, ˜ 1 then the maximal possible e 0 are oriented edges being folded by h meaning that if e, e′ ⊂ H ′ e 1 . Repeat initial segments of e and e that are identified in T+ are identified immediately in H this operation to define an infinite sequence of fold maps h
h
h
3 2 1 ··· H2 −→ H1 −→ H0 −→
˜ i for integers i ≥ 1. By e i → T+ , such that fi−1 = fi ◦ h with Hj ∈ Xbr , and maps fi : H b induction fi is a Λ− isometry and so Hj ∈ Aφ . By construction, any finite segment of this infinite sequence is a sequence of complete Stallings folds in the sense of Section 2.3. Applying Lemma 2.6 it follows that the combinatorial length of the composition hi ◦ · · · ◦ h1 : H0 → Hi goes to infinity as i → +∞. Lemma 6.6 therefore implies that the Hi ’s eventually leave every compact subset of Aφ . Proposition 6.4 implies that there is a compact fundamental domain for the action of φ on Aφ , and the source–sink action of φ on compact subsets of Xr [BFH97] then implies that the closure of Aφ in X r is Aφ ∪ {T− , T+ }. The sequence Hi therefore accumulates on {T− , T+ } as i → +∞. But Len(Hi ) is monotonically decreasing, and so Hi converges to T+ . By interpolating the fold maps hi we produce a fold ray that begins at H and converges to T+ . By Lemma 6.6 there is Λ− isometry from some train track G to H. Factoring this as a sequence of folds and interpolating each fold produces a fold ray that begins with G,
82
passes through H and converges to T+ . Finally, preconcatenating with a periodic split ray that ends at G produces the desired fold line γ. ♦ Example: A fold ray with length 6→ 0. To construct such an example observe that you can do a lot of folds without driving the length to zero, for example by only allowing folds supported in a proper subgraph. For a cheap example, consider any fold ray γ : [0, ∞) → Xr , with Gs representing γ(s). Choose x ∈ G0 . For each s ≥ 0, let G′s be a marked graph of rank r + 1 obtained from Gs by attaching a loop of length 1 based at the point gs0 (x) ∈ Gs a loop of length 1. This yields a fold ray ρ′ : [0, ∞) → Xr+1 with ρ′ (s) represented by G′s , but in the positive direction Len(G′s ) converges to 1, not to zero. Although we shall prove below that every fold ray in outer space converges to a unique limit point in the boundary of outer space, this example shows that the limiting Fr -tree may not have dense branch points. Example: A nonconvergent split ray. Let ρ : (−∞, 0] → Xr be a split ray in rank r outer space Xr that converges to a tree S ∈ ∂Xr , and let σ : (−∞, 0] → Xr′ be a split ray in rank r′ outer space Xr′ that converges to T ∈ ∂Xr′ . Choose connections: a rank r marked graph Gt representing ρ(t) and connection maps gst : Gt → Gs ; and a rank r′ marked graph Ht representing σ(t) and connection maps hst : Ht → Hs . Choose base points xs ∈ Gs so that gst (xt ) = xs and yt ∈ Ht so that hst (yt ) = ys . Identify Fr = π1 (Gs , xs ) and Fr′ = π1 (Ht , yt ). Let Gs ∗ Ht denote the graph obtained from the disjoint union of Gs and Ht by identifying xs to yt to form a base point xyst , and so we have identified π1 (Gs ∗ Ht , xyst ) = Fr+r′ , thus making Gs ∗ Ht into a rank r + r′ marked graph. Consider the third quadrant (−∞, 0] × (−∞, 0] in s, t-space. Any continuous path starting from (s, t) = (0, 0) for which s + t decreases strictly monotonically to −∞ defines a split ray in Xr+r′ , using ρ to split the G factor and σ to split the H factor. Holding t = 0 and letting s → −∞, we obtain a split ray in Xr+r′ converging in X r+r′ to a tree Sˆ on which the Fr′ factor of Fr+r′ = Fr ∗ Fr′ has a fixed point, and so that the Fr action on Sˆ is isometrically conjugate to S. Similarly, holding s = 0 and letting t → −∞, we obtain a split ray in Xr+r′ converging in X r+r′ to a tree Tˆ on which Fr acts with a fixed point, and Fr′ acts isometrically conjugate to T . Now construct a split ray as follows: start from G0 ∗ H0 ; hold ˆ next, t constant and let s decrease, splitting the G factor, so as to approach very close to S; hold s constant and let t decrease, splitting the H factor, so as to approach very close to Tˆ; alternate back and forth in this manner, coming ever closer to Sˆ during the times when G is split, and ever closer to Tˆ when H is split. The result is a split ray in Xr+r′ that accumulates on both Sˆ and Tˆ.
7.2
Characterizing fold lines
We have shown in Proposition 7.1 that Definition (2) of the axis bundle of a fully irreducible φ ∈ Out(Fr ) implies Definition (1): every weak train track for φ lies on some fold line going from T− to T+ . The main result of this section is the converse: each point on each fold line going from T− to T+ is a weak train track for φ, thus proving that Definition (1) implies Definition (2).
83
Theorem 7.2. Given a fully irreducible φ ∈ Out(Fr ) with attracting and repelling fixed points T+ , T− ∈ ∂Xr , if γ : I → Xr is a normalized fold line such that lim γ(t) = T+
t→I+
and
lim γ(t) = T−
t→I−
in
Xr
then I = R and for each t ∈ R, γ(t) is a weak train track for φ. Since Len(γ(t)) = e−t it immediately follows that: Corollary 7.3. The map Len : γ(R) → (0, ∞) is a homeomorphism.
♦
The proof of Theorem 7.2 takes up the rest of Section 7. First we set up the proof, breaking it into three steps, which are carried out in the following subsections. Choose connection data for γ: for each t ∈ I a marked graph Gt of length e−t representing γ(t), and for each for s < t ∈ I an edge isometry gts : Gs → Gt , so that gts ◦ gsr = gtr . Let Tt denote the universal cover of Gt for each t ∈ I. Choose a lift fts : Ts → Tt of each gts , so that fts ◦ fsr = ftr . We must prove that for each t ∈ I there is a map ft+ : Tt → T+ which is injective when restricted to each leaf of Λ− realized in Tt . Along the way we also prove that I = R. There are three steps to the proof. Step 1: Direct limits of fold rays (Section 7.3). Recall the discussion of direct limits in Section 2.5. Recall also Theorem 2.13 which says that from a train track representative 1 i g : G → G of φ one obtains a sequence of marked graphs Gi = λ(φ) i G · φ whose universal covers fit into a direct sequence with direct limit T+ . In other words, T+ is the direct limit of any φ-periodic fold ray. In this section we prove the generalization that for any fold ray which converges to T+ , this convergence is manifested as a direct limit. Regarding the fold ray as a direct system, we first prove that its direct limit is an Fr -tree, and then that Tt converges in Xbr to its direct limit. Using the hypothesis that Tt converges to T+ , we may identify T+ with the direct limit. We obtain direct limit maps ft+ : Tt → T+ satisfying the property that ft+ ◦ fts = fs+ for all s < t. These maps ft+ will be edge isometries. Then, using properties of T+ , we will prove that I+ = +∞. Step 2: Legal laminations of split rays (Section 7.4). Using the fact that Tt converges to T− in X r at t → I− we first prove that the injectivity radius of Gt diverges to +∞ as t → I− . This immediately implies that Len(Gt ) → +∞ and therefore I− = −∞. Next we focus on the “legal lamination”, a sublamination of the geodesic lamination of Fr , defined as the set of all leaves ℓ of the geodesic lamination such that for all r < s in I, the realization of ℓ in Gr is legal for the map fsr . We prove that the legal lamination is nonempty, and so by the Hausdorff maximal principle has a minimal nonempty sublamination L. We then prove that this sublamination has zero length in T− in the sense of [BFH97], and quoting a result of [BFH97] (Lemma 2.19 above) it follows that L = Λ− . Step 3: Fold lines through weak train tracks (Section 7.5). Putting Steps 1 and 2 together, we prove that the realization of Λ− in each Gt , when lifted to Tt , is legal for the map ft+ , and so Gt is a weak train track. 84
7.3
Direct limits of fold rays
In this section we carry out Step 1 in the proof of Theorem 7.2, that a fold ray converging to T+ may be regarded as a direct system with direct limit T+ . Consider a normalized fold path ρ : I → Xr which is noncompact in the positive direction. We shall make no other assumptions on ρ until near the end of the argument when we bring in the hypothesis that ρ(s) → T+ as s → I+ . Choose a normalized connection for ρ: marked graphs Gt representing ρ(t) for each t ∈ I so that Len(Gt ) = e−t ; and edge isometries gts : Gs → Gt so that gts ◦ gsr = gtr . Denote the et . We can lift the map gts to an Fr equivariant map fts : Ts → Tt universal covers as Tt = G so that fts ◦ fsr = ftr . The maps fts are all distance nonincreasing edge isometric maps, and so we have a direct system (Ti )i∈I , (fts : Ts → Tt )s 0 and a+ b = d(ξ, η) then there exists ζ ∈ T∞ such that d(ξ, ζ) = a and d(ζ, η) = b. Fix s0 ∈ I, and choose xs0 , ys0 ∈ Ts0 such that f0∞ (x0 ) = ξ and f0∞ (y0 ) = η. For each s ≥ s0 let xs = fs0 (xs0 ) and ys = fs0 (ys0 ). For all s0 ≤ s < t ∈ I we have fts (xs ) = xt , fts (ys ) = yt , fs∞ (xs ) = ξ and fs∞ (ys ) = η. For each s ≥ s0 choose v s ∈ xs0 ys0 so that fs,s0 (v s ) ∈ xs ys and so that d(xs , fs,s0 (v s )) ≥ a and d(fs,s0 (v s ), ys ) ≥ b; this is possible because xs ys ⊂ fs,s0 (xs0 ys0 ) and because d(xs , ys ) ≥ d(ξ, η) = a + b. Some subsequence of the sequence v s converges to a point zs0 ∈ xs0 ys0 . Set zs = fs,s0 (zs0 ) and ζ = fs∞ (zs0 ) = fs∞ (zs ), for s ≥ s0 . We shall show 0 that ζ is the desired point. Observe that for all s ≥ s0 we have d(xs , zs ) ≥ a. Indeed, if there existed s ≥ s0 such ′ that d(xs , zs ) < a then for all sufficiently large s′ > s we would have d(xs , fs,s0 (v s )) < a ′ which would imply that d(xs′ , fs′ ,s0 (v s )) < a, a contradiction. By a symmetric argument we have d(zs , ys ) ≥ b. Taking infimums as s → ∞, it follows that d(ξ, ζ) ≥ a and d(ζ, η) ≥ b. Next observe that d(xs , zs ) + d(zs , ys ) = d(xs , ys ), and taking infimums as s → ∞ we obtain d(ξ, ζ)+d(ζ, η) = d(ξ, η) = a+b. Combined with the above, it follows that d(ξ, ζ) = a and d(ζ, η) = b.
85
T∞ is nontrivial and minimal. Given s < t ∈ I ∪ {∞}, and given x ∈ Ts , recall the induced map Dfts taking directions of Ts at x to directions of Tt at fts (x), and recall that two directions at x are in the same gate relative to fts if they are identified under Dfts . Lemma 7.4. For all sufficiently large s, each point in Ts has ≥ 2 gates relative to fs∞ . Proof. If d1 and d2 are identified by Dfs∞ then they are identified by Dfts for some finite t. It therefore suffices to prove that if s is sufficiently large then each point in Ts has at least two gates relative to fts for all finite t > s. The property of being in the same gate is invariant under the action of Fr so it suffices to prove that if s is sufficiently large then for each x ∈ Gs and t > s there are directions d1 and d2 at x that are not identified by Dgts . Choose a ∈ I. For each x ∈ Ga consider the following property: (∗) There exists c ≥ a such that if c ≤ s ≤ t then gsa (x) has ≥ 2 gates relative to gts . If x is not a vertex of Ga then Property (∗) is true with c = a, as a consequence of Property (1) in the definition of a fold path. The set of vertices being finite, to complete the proof it remains to verify (∗) for each vertex. If x is a vertex, and if there exists t ≥ a such that −1 gta (gta (x)) contains a point that is not a vertex of Ga , then Property (∗) is true with c = t. −1 Consider the set Va of vertices x of Ga such that for all t ≥ a, the set gta (gta (x)) consists solely of vertices. As t increases the sets Vt = gta (Va ) have nonincreasing cardinality, and so for some c ≥ a the cardinality of Vt is constant for t ≥ c. Suppose that c ≤ s ≤ t, x ∈ Va −1 and consider y = gca (x) ∈ Vc . We have shown that gtc (gtc (y)) = {y}. Since gtc maps Gc onto Gt , it maps a neighborhood of y in Gc onto a neighborhood of gtc (y) in Gt , and so there must be at least two directions at y that are not identified by Dgtc . Property (2) in the definition of a fold path implies that the Dgsc images of these two directions are not identified by Dgts . ♦ For the remainder of the proof we fix s satisfying Lemma 7.4 and we focus on fs∞ . Given x ∈ Ts , a pair of directions at x form a legal turn if they are not in the same gate relative to fs∞ . A bi-infinite geodesic ℓ ⊂ Ts is legal if, for every x ∈ ℓ, the turn made by ℓ at x is legal; equivalently, the restriction of fs∞ to ℓ is an isometric embedding ℓ → T∞ . By Lemma 7.4, every direction at every point x ∈ Ts forms a legal turn with some other direction at x. This implies that every point in Ts lies on some bi-infinite legal geodesic: starting with any edge in Ts , we may extend the edge at each endpoint legally, and continuing to extend by induction we obtain a bi-infinite legal geodesic. To prove that the action of Fr on T∞ is nontrivial and minimal, consider ξ ∈ T∞ , and choose x ∈ Ts such that fs∞ (x) = ξ. Let ℓ ⊂ Ts be a bi-infinite legal geodesic passing through x (in this context we do not assume ℓ is a leaf of Λ− ). Let ℓ− , ℓ+ be the components of ℓ − x. Since there are only finitely many Fr -orbits of oriented edges in Ts , the ray ℓ− must pass over two oriented edges e 6= e′ ⊂ Ts in the same Fr orbit. Let γ− ∈ Fr take e to e′ . Construct a γ− -invariant line by gluing together translates of the segment of ℓ connecting e to e′ . This line must be the axis As (γ− ) of γ− in Ts , and the interior of the segment of ℓ from e to e′ contains a fundamental domain for As (γ− ), implying that As (γ− ) is legal. Doing the same in the other direction, we obtain γ+ ∈ Fr whose axis As (γ+ ) is legal. It follows that γ± are hyperbolic on T∞ and their axes A∞ (γ± ) are the isometric images under fs∞ of the axes As (γ± ); in particular, we now know that the action of Fr on T∞ is nontrivial. By a standard exercise in group actions on R-trees, the axes A∞ (γ± ) are contained in the 86
minimal subtree of T∞ . By construction, there is a segment of Ts passing through x with endpoints respectively on As (γ± ), and so there is a segment of T∞ passing through ξ with endpoints respectively on A∞ (γ± ), implying that ξ is contained in the minimal subtree of T∞ . Since ξ ∈ T∞ is arbitrary, we proved that the action of Fr on T∞ is minimal. Convergence in RC . Next we show that Ts converges to T∞ in RC , that is, for each γ ∈ Fr the translation length of γ acting on Ts converges to the translation length of γ acting on T∞ , as s → I+ . Consider the axis As (γ) ⊂ Ts . For s < t we have fts (As (γ)) ⊃ At (γ). We may therefore choose fundamental domains [xs , ys ] ⊂ As (γ) for s ∈ I so that fts (xs ) = xt , fts (ys ) = yt . Letting ξ = fs∞ (xs ), η = fs∞ (ys ), it follows that d(ξ, η) equals the infimum of the translation lengths. But if ξ 6= η then clearly ∪i∈Z γ i [ξ, η] for i ∈ Z is an axis for γ with translation distance d(ξ, η); whereas if ξ = η then γ(ξ) = η and so γ has translation distance zero. Remark. So far we have not used the hypothesis that Ts converges to T+φ . Our arguments therefore give a completely general proof that every fold ray converges in X r to a unique point T∞ ∈ ∂Xr , and that limit may be regarded as a direct limit. We described earlier an example of a fold ray for which length did not converge to zero. For this fold ray it easily follows that limiting Fr -tree T∞ has nonzero “colength”, meaning that there is a positive lower bound to the length of any finite subtree τ ⊂ T∞ whose translates under Fr cover T∞ . This implies in turn that T∞ has a nondegenerate arc whose interior is an open subset of T∞ , and so the action of Fr on T∞ is not transitive. The action of Fr on T∞ therefore has a nontrivial decomposition, in the sense of [Lev94], as a finite graph of actions with dense branch points. Identification of T∞ with T+ . We have shown that Ts converges to T∞ in RC as s → I+ . Projectivizing, it follows that (the projective class of) T∞ is a point in X r , and that Ts converges to T∞ in X r as s → I+ . But now we use the hypothesis that Ts converges to T+ as s → I+ , to conclude that T∞ equals T+ in Xr , in other words, there exists an Fr -equivariant homothety T∞ ≈ T+ . By scaling the metric on T+ we may conclude that this homothety is an isometry. We thus obtain edge-isometric morphisms fs+ = fs∞ : Ts → T+ = T∞ for s ∈ I, with + fs = ft+ ◦ fts for s < t in I. I+ equals +∞. From the definition of a normalized fold line combined with the definition of the direct limit, it follows that e−I+ equals the limit of the lengths of the marked graphs Gs = Ts /Fr as s → I+ . We show that this limit equals zero. If not, then there exists an open subarc as ⊂ Gs for some s, so that for all t ≥ s, if x ∈ as , y ∈ Gs , and gts (x) = gts (y) then x = y. Letting As ⊂ Ts be a connected lift of as , it follows that for all t ≥ s, if x ∈ As , y ∈ Ts , and fts (x) = fts (y) then x = y. The open arc As therefore embeds as an open subarc of T+ under the direct limit map fs+ : Ts → T+ . However, each orbit of the action of Fr on T+ is dense and so no such open arc exists in T+ . This completes Step 1.
87
7.4
Legal laminations of split rays
In this section we carry out Step 2 in the proof of Theorem 7.2: for any fold ray that converges to T− , the lamination Λ− arises naturally as the “legal lamination” of that fold ray. One can think of this as a generalization of Lemma 2.17, in which it was shown that Λ− is the unique minimal lamination whose leaves are legal with respect to any train track representative of φ. We review the notation: ρ : I → Xr is a normalized fold path assumed to be noncompact in the negative direction; there is a normalized connection for ρ consisting of marked graphs Gt representing ρ(t) for each t ∈ I so that Len(Gt ) = e−t , and edge isometric morphisms et , and lift the gts : Gs → Gt so that gts ◦ gsr = gtr . Denote the universal covers as Tt = G map gts to an Fr equivariant map fts : Ts → Tt of gts so that fts ◦ fsr = ftr . Injectivity radius. Define the injectivity radius of Gt , denoted inj(Gt ), to be the supremum of ǫ > 0 such that for all x ∈ Gt , the ǫ-neighborhood Nǫ (x) is a tree and the inclusion map Nǫ (x) ֒→ Gt is an isometry with respect to the path length metric on Nǫ (x). Note that inj(Gt ) equals half the length of the shortest closed curve on Gt . For the remainder of this section, choose µt > 0 for each t ∈ I so that lim µt Tt = T−
t→I−
Proposition 7.5. As t → I− we have: +∞; and (4) Len(Gt ) → +∞.
b in X r
(1) Len(µt Gt ) → 0;
(2) µt → 0;
(3) inj(Gt ) →
Proof. Note that the Fr -tree T− has an arbitrarily short fundamental domain, meaning that for each ǫ > 0 there exists a point p ∈ T− and a free basis g1 , . . . , gr of Fr such that the distances dT− (p, fi · p) sum to less than ǫ. This is true on general principles because T− has dense branch points, and it can be seen very easily in this case by starting with arbitrary p, g1 , . . . , gr , iterating the action on T− of any automorphism representing φ, and using the fact that φ contracts distance on T− by a factor of λ(φ−1 ) > 1. Since µt Tt converges to T− as t → I− in the Gromov–Hausdorff topology on closed subsets of T− , it follows that for any ǫ > 0, if t is sufficiently close to I− then there exists p ∈ Tt and a free basis g1 , . . . , gr of Fr such that the distances dµt Tt (p, fi · p) sum to less than ǫ. From this it follows immediately that Len(µt Gt ) → 0 as i → I− . But Len(Gs ) ≥ Len(Gt ) for s < t ∈ I, and so limt→I− Len(Gt ) exists and is > 0, possibly +∞. It follows that lim µt = 0. t→I−
Suppose that inj(Gt ) 6→ +∞ as t → I− . Since inj(Gt ) is a nonincreasing function of t, it follows that inj(Gt ) converges to some R > 0 as t → I− . Recall that C is the set of nontrivial conjugacy classes in Fr . Let Ct = {C ∈ C Len(C; Gt ) ≤ 2R}, so Ct is finite and nonempty for all t ∈ I. If s < t ∈ I then Cs ⊂ Ct , because gts is distance nonincreasing. It follows that Ct stabilizes for t sufficiently close to I− , say t ≤ t0 . Let γ be the minimal length closed curve in Gt0 and let C ∈ Ct0 be the conjugacy class of γ, so Len(C; Gt ) ≤ 2R for all t ∈ I. Note that γ is simple and so each representative of C is a generator of a free factor of Fr .
88
Applying Theorem 5.4 of [BFH97], it follows that Len(C; T− ) > 0. On the other hand, Len(C; T− ) = lim Len(C; µt Tt ) = lim µt Len(C; Tt ) t→I−
t→I−
≤ lim µt 2R = 0 t→I−
a contradiction. Since inj(Gt ) → +∞ as t → I− , it follows as well that Len(Gt ) → +∞.
♦
Combining Proposition 7.5 with the normalization condition on the fold ray, we have proved one of the pieces of Theorem 7.2: Proposition 7.6. I− = −∞.
♦
Remark. We have proved that limt→I− Len(Gt ) = +∞ via a proof that inj(Gt ) → +∞, which itself uses the hypothesis that the split ray converges to T− . But inj(Gt ) 6→ +∞ for an arbitrary normalized split ray, as shown by the example described earlier of a split ray in which a certain positive rank subgraph denoted G0 does not change, as the length of the rest of the subgraph goes to +∞. Nevertheless, here is a proof that limt→I− Len(Gt ) = +∞ which is valid for an arbitrary split ray. If not then by monotonicity we have limt→I− Len(Gt ) = L < +∞ and Len(Gt ) ≤ L for all t ∈ I. Using this we show for each C ∈ C that Len(C; Gt ) an upper bound independent of t, for then by monotonicity the translation length function of Gt has a limit in RC whose coordinates are bounded away from zero, and by Lemma 2.4 this corresponds to a point in Xr , contradicting properness of split rays. To bound Len(C; Gt ), fix c ∈ I, and for t ≤ c ∈ I consider a segment α ⊂ Tt whose endpoints are identified by fct . We have Len(Gt ) − Len(Gc ) ≥ 21 Len(α), and so Len(α) ≤ 2(L − Len(Gc )). So if γ ∈ Fr represents C, and if the axis of γ in Tc has K vertices in a fundamental domain, it follows that Len(C; Gt ) = Len(γ; Tt ) ≤ Len(γ; Tc ) + 2K(L − Len(Gc )). A minimal legal lamination. Consider an element ℓ of the geodesic leaf space GFr . Let ℓt denote the realization of ℓ in Tt . We say that ℓ is a legal leaf if for all s < t ∈ I the restriction of fts : Ts → Tt to the geodesic ℓs ⊂ Ts is an embedding, and hence is an isometry onto ℓt . The set of legal leaves is a closed, Fr -equivariant subset of GFr , called f For each t ∈ I the realization of LL f in Tt forms an the legal lamination and denoted LL. f f f t . Let Fr -equivariant sublamination LLt ⊂ GL Tt such that fts pushes LLs forward to LL f f LLt ⊂ GL Gt be the downstairs image of LLt , and so gts pushes LLs forward to LLt . f is not empty. Fact 7.7. LL
Proof. Let Es be the longest edge of Gs and so applying Proposition 7.6 it follows that 1 Len(Gs ) → +∞ as s → −∞. Fixing t ∈ I, for each s < t the path fts Es Len(Es ) ≥ 3r−3 is a legal path in Gt of length equal to Len(Es ), by item (1) in the definition of foldline. Parameterizing this path by a symmetric interval [− Len(Es )/2, Len(Es )/2] and extending the domain to R so that the map is constant on each ray complementary to this interval, we may then pass to a subsequence converging in the compact open topology to a geodesic immersion ℓt : R → Gt . Since ℓt is a union of subsegments of paths fts Es , it follows by 89
item (1) in the definition of foldline that ℓt is legal with respect to all maps fut : Gt → Gu , u ≥ t. Fixing a sequence t1 > t2 > · · · diverging to −∞, by a diagonalization argument we obtain ℓti so that gt1 ti (ℓti ) = ℓt1 . Lifting these geodesics to the universal covers Tti we therefore obtain a legal leaf ℓ. ♦ f has a minimal We now invoke the Hausdorff maximal principle to conclude that LL e This is seen more clearly in nonempty, closed, Fr -invariant subset which we denote L. one of the compact spaces LLt ⊂ GL Gt for any fixed t, because the intersection of any nested sequence of nonempty closed sublaminations of LLt is clearly a nonempty closed sublamination. The realization of Le in GL Tt is denoted Let and the image downstairs in GL Gt is denoted Lt . Recall that Lt is said to be quasiperiodic if for each compact leaf segment α ⊂ ℓ there exists k such that for each compact leaf segment α′ , if Len(α′ ) > k then α′ has a subsegment whose corresponding path in Gt is identical with the path corresponding to α. We list here the properties of the sublaminations Lt : Lemma 7.8. For each t ∈ I, Lt is a nonempty, minimal sublamination of GL Gt , Lt is quasiperiodic, each leaf of Lt is aperiodic, and for each s < t ∈ I, gts maps Ls to Lt inducing a bijection of leaves to leaves and taking each leaf to its corresponding leaf by an isometry. Proof. For any compact lamination minimality implies quasiperiodicity; for convenience we prove this in the present context. Suppose that Lt is not quasiperiodic. Fix a leaf segment α ⊂ Lt that defies the definition of quasiperiodicity. There is a sequence of leaf segments α′n with Len(α′n ) → +∞ so that α′n does not traverse α (we are here confusing a leaf segment with its corresponding path in Gt ). Passing to a limit, we obtain a nonempty sublamination of Lt in which no leaf traverses the segment α, and so this is a sublamination is proper, contradicting minimality. To prove that no leaf is periodic, if ℓt ⊂ Lt is a periodic leaf then the corresponding leaf ℓs ⊂ Ls is periodic for all s < t and Len(ℓs ) = Len(ℓt ). But this implies that inj(Gs ) ≤ 1 ♦ 2 Len(ℓt ) for s < t, contradicting Proposition 7.5 (3). Remark. In an earlier draft we gave a direct but intricate construction which produces sublaminations Lt ⊂ GL Gt satisfying Lemma 7.8 without invoking the Hausdorff Maximal Principle, by studying the inverse limit as t → −∞ of the graphs Gt in the category of compact singular laminations. A careful desingularization process produces the required sublaminations Lt . Identifying L with Λ− . To complete Step 2, we shall show that L has length zero in T− . By applying Lemma 2.19 it will follow that L = Λ− . We must prove that for any marked graph G ∈ Xˆr and every Fr -equivariant morphism e → T− there exists C ≥ 0 such that for every leaf ℓ ⊂ G e of LeG and for any x, y ∈ ℓ, h: G dT− (hx, hy) ≤ C. To prove this it suffices to check it for some G ∈ Xˆr and some Fr e → T ; this is because there exists an Fr -equivariant morphism equivariant morphism h : G ′ e e between G, G for any two G, G′ ∈ Xr , and because any two Fr -equivariant morphisms 90
e → T are Fr -equivariantly homotopic. We will do this for G = G0 and an arbitrarily G e0 → T0 → T− such that φ− is a homothety chosen Fr -equivariant continuous map φ− : G when restricted to each edge of T0 . Continuing with the notation established earlier in Step 2, denote Ts∗ = µs Ts , and recall b as s → −∞; when the metric is not important we will use that T ∗ converges to T in X s
−
r
Ts or Ts∗ without discrimination. Let Ls denote L realized in Gs , and let Les denote the realization in Ts or Ts∗ . We want to connect the property “L has length zero in T− ” to properties of Les for s close to −∞. We claim that there are Fr -equivariant continuous maps φs : T0 → Ts∗ defined for s ≤ 0 satisfying the following property:
Uniform Length: For any finite geodesic segment γ0 ⊂ T0 , denoting γs# = (φs )# (γ0 ) ⊂ # # Ts∗ and γ− = (φ− )# (γ0 ) ⊂ T− , we have Len(γs# ; Ts∗ ) → Len(γ− ; T− ) as s → −∞. To prove this claim, fix a base point in T0 , for R > 0 let BR ⊂ T0 denote the set of vertices of T0 that are within distance R of the base point, let GR denote the set of elements of Fr that move one point of BR to another, and consider the set φ− (BR ) ⊂ T− . For each R > 0 and ǫ > 0, if s is sufficiently large then there exists a GR -equivariant ǫ-approximation from φ− (BR ) to a subset of Ts∗ , and by postcomposing φ− with this ǫ-approximation we ∗ obtain a GR -equivariant map φR,ǫ s : BR 7→ Ts . We may choose functions R(s) → +∞ and R(s),ǫ(s) ǫ(s) → 0 so that φs exists for all s sufficiently close to −∞. We may then extend R(s),ǫ(s) in a unique manner to an Fr -equivariant map φs : T0 → Ts∗ which is homothetic φs on each edge. The Uniform Length property for finite geodesic edge paths γ0 ⊂ T0 follows immediately because the endpoints of γ0 lie on BR for sufficiently large R. For an arbitrary geodesic segment of T0 , which can be written as a finite geodesic edge path with partial edges concatenated at the beginning and at the end, the Uniform Length property follows easily. As a consequence of the Uniform Length property, observe that there are constants L > 0, s0 ∈ I such that the map φ− and each of the maps φs for s ≤ s0 are L-lipschitz: just apply the Uniform Length property to one edge in each of the finitely many edge orbits of T0 . Combining this observation with the bounded cancellation lemma 2.5, there is a constant B > 0 such that: Uniform Bounded Cancellation: The map φ− and each of the maps φs for s ≤ s0 have bounded cancellation constant ≤ B. The constants s0 and B as above shall be fixed for the rest of the argument. Suppose then that L does not have length zero in T− . It follows that there is a finite # leaf segment γ0 of Le0 such that Len(γ− ; T− ) > 6B. Applying the Uniform Length property above, for s sufficiently close to −∞ it follows that Len(γs# ; Ts∗ ) > 5B, let’s say for all s ≤ s1 where s1 ≤ s0 . Since L is minimal there exists a constant K such that for every finite leaf segment β0 of Le0 , if Len(β0 ; T0 ) ≥ K then β0 contains a translate of γ0 . By applying Uniform Bounded Cancellation it follows that Len(βs# ; Ts∗ ) > 3B for all s ≤ s1 and all leaf segments β0 of Le0 with Len(β0 ; T0 ) ≥ K. We will show that this leads to a contradiction. Every geodesic segment in Gs of length equal to inj(Gs ) is embedded and passes over at most 2r − 2 vertices, and therefore intersects some edge in a segment of length at least Cs = inj(Gs )/(2r − 3). It follows that every geodesic segment of length Cs′ = inj(Gs ) + 2 Len(Gs ) 91
contains an entire edge of length at least Cs . There are at most 2r − 1 edges in Gs and so every geodesic segment of length Cs′′ = [2(2r − 1) + 1]Cs′ goes at least three times over some edge whose length is at least Cs . As s → −∞ we have inj(Gs ) → +∞ and so Cs → +∞. As s → −∞ we also have µs Len(Gs ) → 0, and since Cs′′ is bounded above by a constant times Len(Gs ) it follows that µs Cs′′ → 0 as s → −∞. es be a leaf of Les covering a leaf λs of Ls . There is a subpath α es of length at Let λ es of λ e1 α e2 α e3 where E e1 , E e2 , E e3 are lifts of the same edge E most Cs′′ that decomposes as α es = E e′ E e′′ E of Gs with length at least Cs . If every crossing of E by the leaf λs was immediately followed by α′′ E then λs would be the periodic leaf (α′′ E)∞ . Since λs is not periodic it follows that e′ of λ es that contains E e2 but not E e2 α e3 . Similarly, there is there is a covering translate λ e′′ E s ′′ ′e e es ∩ λ e′ ∩ λ e′′ e e a covering translate λs that contains E2 but not E1 α e E2 . Thus the arc σ es = λ s s satisfies Cs ≤ Len(e σs ; Ts ) ≤ Cs′′ and so Len(e σs ; Ts∗ ) ≤ µs Cs′′ → 0
as s → −∞
Let β(s) = f0s (e σs ). For s sufficiently close to −∞ we have Len(β(s); T0 ) = Len(e σs ; Ts ) ≥ Cs ≥ K and so
# Len (β(s))s ; Ts∗ > 3B
es , λ e′ , λ e′′ , when mapped from Ts to T0 via f0s , give On the other hand, the geodesics λ s s e′′ , and by construction we have e0 , λ e′ , λ geodesics in T0 denoted λ 0 0 e0 ∩ λ e′ ∩ λ e′′ β(s) ⊂ λ 0 0
Applying Uniform Bounded Cancellation to the map φs it follows that if s ≤ s0 then e e′ e′′ (β(s))# es , and so s = (φs )# (β(s)) is contained in the B-neighborhood of λs ∩ λs ∩ λs = σ for s sufficiently close to −∞ we have ∗ σs ; Ts∗ ) < 3B Len (β(s))# s ; Ts < 2B + Len(e which is a contradiction. This contradiction shows that L must have length zero in T− . Since L is minimal, Lemma 2.19 shows that L = Λ− , completing Step 2.
7.5
Weak train tracks on fold lines
Continuing with the notations of Sections 7.3 and 7.4, we now carry out Step 3, completing the proof of Theorem 7.2. By Section 7.3 the direct limit of the system of Fr -equivariant maps fst : Tt → Ts is T+ , and we have direct limit maps fs : Ts → T+ , each of which is Fr -equivariant and an isometry on each edge. Fix a leaf ℓ of Λ− and s ∈ R, and let ℓs ⊂ Ts denote the realization of ℓ in Ts . We must show that fs ℓs is an isometry from ℓs into T+ , that is, for each x, y ∈ ℓs we must prove dT+ (fs (x), fs (y)) = dTs (x, y). In Section 7.4 we proved that Λ− = L is the unique minimal legal sublamination. The geodesic ℓs is therefore legal, and so fts ℓs is an isometry 92
for each t ≥ s. For any t ≥ s it follows that dTt (fts (x), fts (y)) = dTs (x, y). Since the direct sequence (fts (x))t≥s determines fs (x) ∈ T+ , and the sequence (fts (y))t≥s determines fs (y), it follows that dT+ (fs (x), fs (y)) equals the limit of the constant sequence dTt (fts (x), fts (y)) which equals dTs (x, y), proving that dT+ (fs (x), fs (y)) = dTs (x, y). This proves that each map fs : Gs → T+ is a Λ− isometry, and so Gs is a weak train track.
93
References [BB97]
M. Bestvina and N. Brady, Morse theory and finiteness properties of groups, Invent. Math. 129 (1997), no. 3, 445–470.
[BF]
M. Bestvina and M. Feighn, Outer limits, preprint.
[BFH97] M. Bestvina, M. Feighn, and M. Handel, Laminations, trees, and irreducible automorphisms of free groups, Geom. Funct. Anal. 7 (1997), 215–244. [BFH00]
, The Tits alternative for Out(Fn ). I. Dynamics of exponentially-growing automorphisms., Ann. of Math. 151 (2000), no. 2, 517–623.
[BFH04]
, Solvable subgroups of Out(Fn ) are virtually Abelian, Geometriae Dedicata 104 (2004), 71–96.
[BH92]
M. Bestvina and M. Handel, Train tracks and automorphisms of free groups, Ann. of Math. 135 (1992), 1–51.
[Cla05]
M. Clay, Contractibility of deformation spaces of G-trees, Algebr. Geom. Topol. 5 (2005), 1481–1503 (electronic).
[CM87]
M. Culler and J. W. Morgan, Group actions on R-trees, Proc. London Math. Soc. (3) 55 (1987), no. 3, 571–604.
[Coo87]
D. Cooper, Automorphisms of free groups have finitely generated fixed point sets, J. Algebra 111 (1987), no. 2, 453–456.
[CP94]
M. Coornaert and A. Papadopoulos, Une dichotomie de Hopf pour les flots g´eod´esiques associ´es aux groupes discrets d’isom´etries des arbres, Trans. AMS 343 (1994), no. 2, 883–898.
[CV86]
M. Culler and K. Vogtmann, Moduli of graphs and automorphisms of free groups, Invent. Math. 84 (1986), 571–604.
[FH04]
M. Feighn and H. Handel, Abelian subgroups of Out(Fn ), preprint, 2004.
[FLP+ 79] A. Fathi, F. Laudenbach, V. Poenaru, et al., Travaux de Thurston sur les surfaces, Ast´erisque, vol. 66–67, Soci´et´e Math´ematique de France, 1979. [GdlH91] E. Ghys and P. de la Harpe, Infinite groups as geometric objects (after Gromov), Ergodic theory, symbolic dynamics and hyperbolic spaces (Trieste, 1989) (T. Bedford, M. Keane, and C. Series, eds.), Oxford Univ. Press, 1991, pp. 299–314. [GJLL98] D. Gaboriau, A. Jaeger, G. Levitt, and M. Lustig, An index for counting fixed points of automorphisms of free groups, Duke Math. J. 93 (1998), no. 3, 425–452. [GL05]
V. Guirardel and G. Levitt, The outer space of a free product, preprint, arXiv:math.GR/0501288, 2005.
[HM06a] M. Handel and L. Mosher, The expansion factors of an outer automorphism and its inverse, Trans. AMS (2006), To appear. Preprint, arXiv:math.GR/0410015. 94
[HM06b]
, Parageometric outer automorphisms of free groups, Trans. AMS (2006), To appear. Preprint, arXiv:math.GR/0410018.
[HT85]
M. Handel and W. Thurston, New proofs of some results of Nielsen, Adv. in Math. 56 (1985), no. 2, 173–191.
[Lev94]
G. Levitt, Graphs of actions on R-trees, Comment. Math. Helv. 69 (1994), no. 1, 28–38.
[LL00a]
G. Levitt and M. Lustig, Most automorphisms of a hyperbolic group have very ´ simple dynamics, Ann. Sci. Ecole Norm. Sup. (4) 33 (2000), no. 4, 507–517.
[LL00b]
, Periodic ends, growth rates, H¨ older dynamics for automorphisms of free groups, Comment. Math. Helv. 75 (2000), no. 3, 415–429.
[LL03]
, Irreducible automorphisms of Fn have north-south dynamics on compactified outer space, J. Inst. Math. Jussieu 2 (2003), no. 1, 59–72.
[LL04]
J. Los and M. Lustig, The set of train track representatives of an irreducible free group automorphism is contractible, preprint, December 2004.
[Lus99]
M. Lustig, Structure and conjugacy for automorphisms of free groups, preprint, 1999.
[McC85]
J. McCarthy, A “Tits-alternative” for subgroups of surface mapping class groups, Trans. AMS 291 (1985), no. 2, 582–612.
[MS93]
H. Masur and J. Smillie, Quadratic differentials with prescribed singularities and pseudo-Anosov diffeomorphisms, Comment. Math. Helv. 68 (1993), no. 2, 289– 307.
[Nie86]
J. Nielsen, Investigations in the Topology of Closed Orientable Surfaces, I, Jakob Nielsen: Collected Mathematical Papers, Volume 1, Birkh¨auser, 1986, Translation by John Stillwell of: Untersuchungen zur Topologie der geschlossenen zweiseitigen Fl¨ achen. Acta. Math., 50 (1927), 189–358.
[Pau89]
F. Paulin, The Gromov topology on R-trees, Topology Appl. 32 (1989), no. 3, 197–221.
[Sko]
R. Skora, Deformations of length functions in groups, preprint.
[Sko90]
, Splittings of surfaces, Bull. AMS 23 (1990), 85–90.
[Spa81]
E. Spanier, Algebraic topology, Springer, New York—Berlin, 1981.
[Sta83]
J. Stallings, Topology of finite graphs, Inv. Math. 71 (1983), 551–565.
[Vog02]
K. Vogtmann, Automorphisms of free groups and outer space, Proceedings of the Conference on Geometric and Combinatorial Group Theory, Part I (Haifa, 2000), vol. 94, 2002, pp. 1–31.
[Whi93]
T. White, Fixed points of finite groups of free group automorphisms, Proc. AMS 118 (1993), no. 3, 681–688.
95